A drug-drug interaction prediction method based on RGDA-DDI

Abstract

The present invention relates to the field related to artificial intelligence and drug discovery. Specifically, a drug-drug interaction (DDI) prediction method based on RGDA-DDI is invented to solve the problem that the existing DDI prediction methods are not ideal. The method is divided into a data encoding module, a feature fusion module and a prediction module. The data encoding module is composed of multiple Residual-GAT submodules, each of which is composed of a graph attention layer, a Normalize layer and a SAGPooling layer, and is used to extract multi-scale features of the input drugs; the feature fusion module is composed of two dual-attention mechanism submodules, which are used for multi-scale drug feature fusion; finally, the fused features are input into the prediction module for prediction. This method overcomes the shortcomings of the existing DDI prediction methods, which lack modeling of multi-scale drug features and lack of mining of potential drug pairs (DDP) features. Experimental verification shows that the prediction performance of this method is better than the recently disclosed drug-drug interaction prediction method.

Description

[0001] Technical Field: The present invention relates to the field of artificial intelligence and drug discovery. Specifically, a drug-drug interaction prediction method based on RGDA-DDI is invented. Background Art

[0002] During medication use, two or more drugs are often used simultaneously. This approach has a positive effect on promoting treatment, but it also increases the potential risks associated with the combined use of different drugs. Therefore, drug-drug interaction analysis is an essential task in the drug discovery process. Traditional drug-drug interaction analysis is labor-intensive, requiring extensive biochemical experiments and consuming considerable time. Artificial intelligence-based drug discovery has become a hot research area in recent years. AI technology can analyze the degree of interaction between different drugs, facilitating the rapid identification of optimal drug-drug pairs (DDPs). Therefore, AI technology can significantly shorten drug discovery timelines and reduce repetitive tasks in drug-drug interaction (DDI) prediction, significantly promoting drug discovery efforts.

[0003] 1. Professional terminology

[0004] (1) Deep Learning (DL). Deep learning technology has achieved very brilliant achievements in recent years, especially in natural language processing, image processing and other fields. This technology is developed from multi-layer neural networks. Its essence is to build a machine learning model with a large number of neural network hidden layers, and use massive training data to train and learn more representative features to increase classification accuracy. The main purpose of deep learning is "feature learning". Unlike traditional shallow learning, deep learning often has more hidden layers. Through the feature interaction between each layer, the original feature representation is transformed into a new feature space, and then the loss function and optimization function are used to optimize the training effect through the feature interaction information, thereby optimizing the model.

[0005] (2) Drug Knowledge Graph (DKG). A knowledge graph is a relational network that connects different types of information. It provides structural relationships between multiple entities and unstructured semantic relationships between each entity. In the field of drug discovery, a knowledge graph constructed with drug molecules as entities is called a drug knowledge graph (DKG).

[0006] (3) Graph Attention Networks (GAT). Graph Attention Networks are a combination of Graph Neural Networks (GNN) and attention mechanisms. It is a neural network architecture based on graph structured data. It can use hidden self-attention layers to solve the shortcomings of previous methods based on graph convolution or its approximation. The operation of the self-attention layer can be performed in parallel between all edges, and the calculation of the output features of all nodes can also be performed in parallel, without the need for feature decomposition and similar costly matrix operations. Therefore, compared with graph neural networks, graph attention networks are more efficient. In addition, GAT allows different weights to be assigned to nodes with the same neighbors, thereby achieving an increase in model capacity. Based on the above characteristics, the present invention applies GAT to the field of drug-drug interaction prediction, achieving ideal results, and has certain novelty and creativity.

[0007] 2. Related research progress

[0008] According to relevant literature at home and abroad, good results related to drug-drug interaction prediction have emerged in recent years. For example, LaGAT (Hong et al. 2022) uses the embedded representation of one of the drugs as the query vector to calculate the attention weight, and selects appropriate topological neighbor nodes to obtain the semantic information of the other drug. Finally, the two different drug information are concatenated and input into the prediction module for DDI prediction; KGNN (Lin et al. 2020) extracts drug features through graph neural network GNN and knowledge graph KG, samples and aggregates the neighborhood information of each node from the local receiver of each node, and then predicts DDI; SumGNN (Yu et al. 2021) uses a subgraph summary framework based on self-attention mechanism to generate paths in the subgraph, and adopts multi-channel knowledge and data integration modules for DDI prediction; DSN-DDI (Li et al. 2023) iteratively uses local and global representation learning modules, and learns drug feature information in single drugs (within view) and drug pairs (between views), and finally inputs the extracted information into the prediction module for DDI prediction; BioDKG-DDI (Ren et al. 2022) integrated multiple biochemical features to predict DDI via an attention machine.

[0009] The patent document (publication number: CN 115458044 A) discloses a drug-drug interaction prediction method based on the global structure of the network. The method first extracts data from the database and constructs a biological heterogeneous network. Then, the shortest path is manually selected and the data is annotated. The data is input into a self-supervised learning model for training for DDI prediction. However, this method does not provide the architecture and construction method of the biological heterogeneous network, nor does it provide the model architecture of the self-supervised learning model. Therefore, it is difficult for technicians to construct the model from the content of the invention or the implementation method, and it is impossible to verify the creativity and practicality of the model. The patent document (publication number: CN 115512761 A) discloses a drug-drug interaction prediction framework based on meta-paths. The framework uses the semantic relationship between different biological entities to obtain the characteristic representation of the drug, thereby predicting DDI. However, the framework uses the information fusion mechanism based on meta-paths to obtain drug-protein information, but does not provide a clear explanation of the specific contribution to drug-drug interaction prediction. A patent document (publication number: CN 114093413 A) discloses a drug-drug interaction prediction system and method based on deep learning. This invention uses multiple graph attention networks (GATs) to extract drug information, which is then concatenated and iterated multiple times before being input into a fully connected layer for DDI prediction. However, this invention does not consider the impact of the characteristic relationships between individual drugs and drug pairs (DDPs) on DDI prediction. A patent document (publication number: CN 113643765A) discloses a drug-drug interaction prediction method based on a tensor neural network. This method first converts the acquired drugs into a vector representation of a fixed dimension, then combines the substructures and interactions into a fully connected coefficient matrix, and performs DDI prediction after multiple linear transformations. This method involves a large amount of manual work and feature extraction, and the number of extracted features is limited compared to deep learning-based methods, which can affect the DDI prediction effect.

[0010] Problems with currently available drug-drug interaction prediction methods

[0011] The above methods have certain reference significance for understanding the research progress of drug-drug interaction prediction. However, the existing drug-drug interaction prediction methods also have some limitations, mainly including:

[0012] (1) Most existing methods focus on encoding the input drugs separately, ignoring the interaction between single drugs and drug-drug pairs (DDPs), that is, they do not model the multi-scale drug feature representation.

[0013] (2) Most computer-based drug-drug interaction prediction methods focus on extracting drug features and then simply concatenating the extracted drug features. This method does not take into account the in-depth mining of potential drug features and the impact of efficient feature fusion methods on prediction results, making it difficult to further improve the effect of DDI prediction.

[0014] To address these issues, the present invention discloses a drug-drug interaction prediction method, RGDA-DDI, based on a combination of Residual-GAT and dual-attention. This method utilizes a residual-GAT module to simultaneously learn multi-scale drug feature representations from individual drugs and drug pairs (DDPs). It then fuses these learned features using a dual-attention mechanism, which are then fed into a prediction module for drug-drug interaction prediction. Summary of the Invention

[0015] The present invention aims to solve the problem of low prediction accuracy caused by incomplete feature acquisition and unscientific feature fusion in existing drug-drug interaction prediction methods. A drug-drug interaction prediction method RGDA-DDI based on the combination of Residual-GAT and dual-attention is invented. The method specifically comprises the following steps:

[0016] S1: Data acquisition: Obtain drug information from the drug database, including drug compound structure information and verified drug-drug relationship information, and store the obtained drug data in SMILS format;

[0017] S2: Data preprocessing: converting drug information and drug pair information into matrix form. Verified drug-drug interaction information is marked as positive samples, and unverified drug-drug pairs are marked as negative samples. To facilitate model calculation, the same number of positive samples and negative samples are randomly selected.

[0018] S3: Construct drug knowledge graph DKG and define Where D={d1,d2,…,d n}

[0019] represents a drug data set, Drug d i and d j The set of mutual relations between them. At the same time, the triple (d i ,r ij ,d j) is used to indicate drug d i and d j The type of relationship between i ,d j ∈D,

[0020] S4: Construction of drug-drug interaction prediction model based on RGDA-DDI, please refer to the attached Figure 1 The model is mainly divided into data encoding module, feature fusion module and prediction module; the data encoding module is mainly composed of several Residual-GAT submodules. Please refer to the attached data flow diagram of the Residual-GAT submodule. Figure 2 Each Residual-GAT submodule is composed of a graph attention network (GAT), a Normalize layer, and a SAGPooling layer. Two local residual concatenations and one global residual concatenation are used for data transmission. For the architecture of the SAGPooling layer, please refer to the attached Figure 3 The architecture first consists of a three-layer graph neural network (GNN), followed by residual concatenation, which is then input into the graph pooling layer for feature extraction. Finally, the extracted features are input into the readout layer for feature fusion and output. The feature fusion module is composed of two dual-attention mechanism submodules. Please refer to the architecture of each attention mechanism submodule. Figure 4 , this submodule contains two attention structures, which are used to receive local drug features and global drug features respectively. Each attention structure is composed of vectors Q, K, and V, respectively. Represents query vector query, key vector key and value vector value respectively, N and M represent the length of vector Q, K, V respectively, D represents the dimension of feature matrix, Q and K in each attention structure are respectively input into tanh function with vector K and Q in another attention structure, and the obtained value is multiplied with vector V in this attention mechanism to obtain the DDI feature extracted by the attention structure; finally, the feature vectors calculated by the two attention mechanism modules are input into the prediction module for DDI result calculation; the prediction module is composed of fully connected layers, which selects different activation functions according to different tasks: the activation function used in the binary DDI interaction prediction is Sigmoid function, which can be expressed as in Represents the vector v in the binary classification prediction task i and vj Is there any relationship between them? If so, on the contrary In multi-classification tasks, DDI prediction mainly calculates drug d i and d j There is some type of relationship between ij The probability value of , the activation function used is the Softmax function, which is expressed by the formula in Representative drug d i and d j The type of relationship between probability;

[0021] S5: Use stratified N-fold cross-validation to train the model (N ≥ 2). The data of different categories are randomly divided into N parts with the same percentage, so that the percentage of each category in the training set and the test set is the same. This ensures that there is no case where data of a certain category is in the training set but not in the test set.

[0022] S6: The drug and DDI data in the training data were input into RGDA-DDI for model training. The Adam function was used as the optimization function, and its initial learning rate was set to 0.0005. The binary cross-entropy loss function was used to evaluate the difference between the predicted results and the actual solution. If there were multiple types of DDIs between drugs, the cross-entropy loss function was used to evaluate the difference between the predicted results and the actual solution. The training was stopped by setting a certain number of iterations until the model no longer converged and the trained model was saved.

[0023] S7: Use the test set to test the model prediction effect. If the DDI prediction result is not ideal, continue to adjust the model's hyperparameters and repeat the training process in step 6 until the prediction module outputs a DDI test result that meets the accuracy requirement. Then stop training and obtain a trained DDI prediction model.

[0024] S8: Input any two drugs that meet the format requirements into the trained DDI prediction model. After calculation, the calculation results of the input drug-drug interaction can be obtained.

[0025] The present invention provides a drug-drug interaction prediction method based on RGDA-DDI. This method overcomes the shortcomings of existing DDI prediction methods, which lack explicit modeling of multi-scale drug feature representations and fail to consider deep mining of potential features during drug feature fusion. Experimental validation demonstrates that this method significantly improves predictive performance on publicly available datasets compared to recently published DDI prediction models. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] Attachment Figure 1 : Architecture diagram of drug-drug interaction prediction method based on RGDA-DDI;

[0027] Attachment Figure 2 : Residual-GAT submodule architecture diagram;

[0028] Attachment Figure 3 : SAGPooling architecture diagram;

[0029] Attachment Figure 4 : Dual-attention mechanism feature fusion module architecture diagram. DETAILED DESCRIPTION

[0030] For a better understanding of the purpose, technical solutions and advantages of the present invention, the present invention will be further described below in conjunction with the accompanying drawings and specific example embodiments. Those skilled in the art can easily understand the advantages and effects of the present invention from the content disclosed in this specification, but the present invention is not limited in any form. It should be pointed out that, for those of ordinary skill in the art, without departing from the idea of ​​the present invention, several changes and improvements can also be made, and these all fall within the scope of protection of the present invention. Below in conjunction with the accompanying drawings, some embodiments of the specific examples of the present invention are described in detail. In the absence of conflict, the following embodiments can be extended to all drug-drug interaction predictions.

[0031] According to the present invention, a drug-drug interaction prediction method based on RGDA-DDI is provided. The main structure of this method is shown in the attached Figure 1 , combined with specific examples, the main implementation steps include:

[0032] S1: Data acquisition: Obtain drug information from the drug database, including drug compound structure information and verified drug-drug interaction information. The acquired drug data is stored in SMILS format. In this example, drug information is obtained from the DrugBank database. The drug data is then converted into an RDF graph using a conversion tool (Bio2RDF). The RDF graph is then stored in the form of triples. The SPARQL query tool is then used to extract the (entity 1, relationship, entity 2) triples.

[0033] S2: Data preprocessing: convert drug and drug pair information into matrix representation. Verified drug-drug pairs with interactions are labeled as positive samples, and unverified drug-drug pairs are labeled as negative samples. To facilitate model calculation, an equal number of positive and negative samples are randomly selected. The drug data collected in this example includes 1,925 drug molecules and 56,983 verified drug-drug interaction data (positive samples), as well as an equal number of negative samples.

[0034] S3: Construct drug knowledge graph DKG and define Where D={d1,d2,…,d n} represents the drug data set, Drug d i and d j The set of mutual relations between them. At the same time, the triple (d i ,r ij ,d j ) is used to indicate drug d i and d j The type of relationship between i ,d j ∈D,

[0035] S4: Construction of drug-drug interaction prediction model based on RGDA-DDI, please refer to the attached Figure 1 The model is mainly divided into data encoding module, feature fusion module and prediction module; the data encoding module is mainly composed of several Residual-GAT submodules. Please refer to the attached data flow diagram of the Residual-GAT submodule. Figure 2Each Residual-GAT submodule is composed of a graph attention network (GAT), a Normalize layer, and a SAGPooling layer. Two local residual connections (local residual concatenation) and one global residual connection (global residual concatenation) are used for data transmission between them. The SAGPooling layer architecture is detailed in the appendix. Figure 3 The feature fusion module is composed of a cross-combination of dual-attention mechanisms. Figure 4 , this module consists of two self-attention mechanism submodules, each self-attention can be expressed as: in Represents query vector query, key vector key and value vector value respectively, N and M represent the length of vector Q, K, V respectively, and D represents the dimension of feature matrix. In the feature fusion module, the vectors Q and K in each self-attention mechanism are respectively input into the tanh function with the vectors K and Q in another self-attention mechanism. The obtained value is multiplied with the vector V in the self-attention mechanism to obtain the fused DDI feature, which is finally input into the DDI prediction module for calculation; the prediction module is composed of a fully connected layer, which selects different activation functions according to different tasks: the activation function used in the binary DDI interaction prediction is the Sigmoid function, which can be expressed as in Represents the vector v in the binary classification prediction task i and v j Is there any relationship between them? If so, on the contrary In multi-classification tasks, DDI prediction mainly calculates drug d i and d j There is some type of relationship between ij The probability value of , the activation function used is the Softmax function, which is expressed by the formula in Representative drug d i and d j The type of relationship between probability;

[0036] S5: Set model parameters. In order to achieve the desired prediction effect of the designed model, reasonable hyperparameters need to be set. The specific parameters and parameter values ​​of this example are detailed in Table 1.

[0037] Table 1 Hyperparameters and parameter values ​​used in this example

[0038]

[0039] S6: Use stratified N-fold cross-validation to train the model (N ≥ 2). The data is randomly divided into N parts with the same percentage, so that the percentage of each category in the training set and the test set is the same. This ensures that there is no case where a certain category of data is in the training set but not in the test set.

[0040] The drug and DDI data from the training data were fed into RGDA-DDI for model training. This example used the Adam function for optimization, with an initial learning rate of 0.0005. The binary cross-entropy loss function was used to evaluate the difference between the predicted results and the true solution. If multiple DDI types exist between drugs, the cross-entropy loss function was used to evaluate the difference between the predicted results and the true solution. Training was stopped by setting a set number of iterations until the model no longer converged, and the trained model was saved.

[0041] S7: Use the test set to test the model prediction effect. If the DDI prediction result is not ideal, continue to adjust the model's hyperparameters and repeat the training process in step 6 until the prediction module outputs a DDI test result that meets the accuracy requirement. Then stop training and obtain a trained DDI prediction model.

[0042] S8: Input any two drugs that meet the format requirements into the trained DDI prediction model. After the model calculation, the calculation results of the input drug-drug interaction can be obtained;

[0043] S9: Comparison of experimental results. The evaluation indicators used in this example are accuracy, area under the receiver operating characteristic curve (AUROC), area under the precision / recall curve (AUPR), and the average of precision and recall F1-score. The comparison of the experimental results of the above evaluation indicators on the dataset used in this example is shown in Table 2.

[0044] Table 2. Comparison of experimental results (bold values ​​represent the best results)

[0045]

[0046] Through comparison, it is found that the performance of the method of the present invention is significantly better than existing DDI prediction methods such as LINE, GCN, GAT, Deepwalk, Deep-DDI, Conv-LSTM, LAGAT, KGNN, and MUFFIN.

[0047] The above describes specific embodiments of the present invention. It should be understood that the present invention is not limited to the specific embodiments described above, and those skilled in the art may make various changes or modifications within the scope of the claims, including but not limited to implementing the same program in the form of logic gates, switches, application-specific integrated circuits, programmable logic controllers, and embedded microcontrollers in systems, devices, and modules designed based on the methods provided by the present invention, without affecting the essence of the present invention.

[0048] It should be understood that the flowcharts and block diagrams in the accompanying drawings show the possible architectures, functions and operations of the systems, methods and computer program products according to multiple embodiments of the present invention. Each box in the flowchart or block diagram can represent a module, program segment or part of the code, and these flowcharts or block diagrams can be implemented by a dedicated hardware-based system that performs the specified functions or actions, or by a combination of dedicated hardware and computer instructions. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. The technicians can use different methods to implement the described functions for each specific application, which does not exceed the scope of protection of the present invention.

Claims

1. A drug-drug interaction prediction method based on RGDA-DDI, characterized in that: The following steps are involved: S1: Data acquisition, obtaining drug information from the drug database, including drug compound structure information and verified drug-drug interaction pair information; S2: Data preprocessing: converting drug information and drug pair information into matrix form. Verified drug-drug interaction information is marked as positive samples, and unverified drug-drug pairs are marked as negative samples. To facilitate model calculation, the same number of positive samples and negative samples are randomly selected. S3: Building a drug knowledge graph Where D={d1,d2,…,d n } represents the drug data set, Drug d i and d j The set of mutual relations between them, and define the triples (d i ,r ij ,d j ) is used to indicate drug d i and d j The type of relationship between i ,d j ∈D, S4: Construct a drug-drug interaction prediction model based on RGDA-DDI. The model includes a data encoding module, a feature fusion module, and a prediction module. The data encoding module mainly consists of several Residual-GAT submodules. Each Residual-GAT submodule is composed of a graph attention network (GAT), a Normalize layer, and a SAGPooling layer. Two local residual connections and one global residual connection are used for data transmission between them. SAGPooling consists of three graph neural network (GNN) layers, a graph pooling layer, and a readout layer, which are then combined through residual concatenation connections. S5: Use stratified N-fold cross-validation to train the model, with N ≥ 2. The data is randomly divided into N parts with the same percentage, so that the percentage of each category is the same in the training set and the test set. This ensures that there is no case where a certain category of data is in the training set but not in the test set. S6: The drug and DDI data in the training data were input into RGDA-DDI for model training. The Adam function was used as the optimization function, and its initial learning rate was set to 0.0005. The Binary cross-entropy loss function was used to evaluate the difference between the predicted results and the actual solution. If there were multiple DDI types between drugs, the cross-entropy loss function was used to evaluate the difference between the predicted results and the actual solution. The training was stopped by setting a certain number of iterations until the model no longer converged and the trained model was saved. S7: Use the test set to test the model prediction effect. If the DDI prediction result is not ideal, continue to adjust the model's hyperparameters and repeat the training process of step S6 until the prediction module outputs a DDI test result that meets the accuracy requirement. Stop training and obtain a trained DDI prediction model. S8: Input any two drugs that meet the format requirements into the trained DDI prediction model. After calculation, the calculation results of the input drug-drug interaction can be obtained.

2. The method for predicting drug-drug interactions based on RGDA-DDI according to claim 1, characterized in that: The drug data retrieved from the database are stored in SMILS format.

3. The method for predicting drug-drug interactions based on RGDA-DDI according to claim 1, characterized in that: The RGDA-DDI prediction model includes a data encoding module, a feature fusion module, and a prediction module. The feature fusion module is composed of two dual-attention mechanism submodules. Each attention mechanism submodule contains two attention structures, which are used to receive local drug features and global drug features respectively. Each attention structure is composed of vectors Q, K, and V, where Represents query vector query, key vector key and value vector value respectively, N and M represent the length of vector Q, K, V respectively, D represents the dimension of feature matrix, Q and K in each attention structure are respectively input into tanh function with vector K and Q in another attention structure, and the obtained value is multiplied with vector V in this attention mechanism to obtain the DDI feature extracted by the attention structure; finally, the feature vectors calculated by the two attention mechanism modules are input into the prediction module for DDI result calculation; the prediction module is composed of fully connected layers, which selects different activation functions according to different tasks: the activation function used in the binary DDI interaction prediction is Sigmoid function, which can be expressed as in Represents the vector v in the binary classification prediction task i and v j Is there any relationship between them? If so, on the contrary In multi-classification tasks, DDI prediction mainly calculates drug d i and d j There is some type of relationship between ij The probability value of , the activation function used is the Softmax function, which is expressed by the formula in Representative drug d i and d j The type of relationship between probability.

4. The method for predicting drug-drug interactions based on RGDA-DDI according to claim 1, characterized in that: If the prediction task is a binary classification task, that is, predicting whether there is a relationship between two drugs, choose the Binary cross-entropy loss function; if the prediction task is to predict multiple types of relationships between two drugs and output the probability of belonging to a certain type of drug-drug relationship, choose the Cross-entropy loss function.

5. The method for predicting drug-drug interactions based on RGDA-DDI according to claim 1, characterized in that: This method is used to predict drug-drug interactions.

Images