Drug interaction prediction method and system based on bilinear graph neural network
By constructing a heterogeneous information network and using a bilinear polymerizer, the problems of heterogeneous graph structure fusion and neighbor node relationship capture in drug interaction prediction are solved, and more efficient drug interaction prediction is achieved.
Patent Information
- Application Number
- CN202510230998.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-02-28
- Publication Date
- 2025-07-25
AI Technical Summary
Existing drug interaction prediction methods based on graph neural networks are difficult to effectively fuse heterogeneous graph structures and capture complex nonlinear relationships between neighbor nodes, resulting in insufficient drug interaction prediction performance.
A heterogeneous information network was constructed, using bilinear polymerizers and metapathic information fusion mechanisms, designing attention mechanisms to learn drug representations, and predicting drug interactions through bilinear polymerizers and graph convolution operations.
Improves the accuracy and efficiency of drug interaction prediction, enables more comprehensive capture of the complex semantics of drug-drug interactions, adapts to multimodal data scenarios, and provides partial interpretability.
Smart Images
Figure CN120375971A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of drug interaction prediction, and in particular to a drug interaction prediction method and system based on a bilinear graph neural network. Background Art
[0002] Drug-Drug Interaction Prediction (DDIP) is crucial in drug research and development and pharmacovigilance. The occurrence of DDI can lead to adverse reactions, reduce efficacy, and even endanger the lives of patients. Traditional DDI detection methods rely on time-consuming and costly wet laboratory experiments, which seriously hinder the drug development process. To address this issue, researchers have explored computational DDI prediction methods, especially those leveraging deep learning techniques. Among them, Graph Neural Networks (GNNs) have shown significant advantages in the field of DDI prediction due to their powerful graph data processing capabilities. GNNs can model drugs and their interaction relationships as a graph structure and learn the latent feature representations of drugs to predict potential DDIs. However, existing GNN-based DDI prediction methods still have some limitations. Traditional GNN models usually process the neighbor information of each node independently, ignoring the interactions between neighbor nodes, and thus failing to fully capture the complex biological mechanisms of DDIs. In addition, some methods only focus on drug molecular structure information while ignoring the complex interaction relationships between drugs and proteins, such as drug-target interactions, drug-carrier interactions, drug-enzyme interactions, and drug-transporter interactions. These information are crucial for understanding the occurrence mechanism of DDIs. Therefore, how to effectively integrate multi-source heterogeneous information such as drugs, proteins, and their interaction relationships, and fully consider the interactions between neighbor nodes, has become the key challenge in improving DDI prediction performance. Specifically, the following two key problems need to be solved urgently in this field:
[0003] (1) How to effectively fuse heterogeneous graph structures to more comprehensively capture the complex semantics of DDIs. Most existing GNN models are based on homogeneous graph structures and are difficult to effectively integrate different types of biomedical information. At the same time, traditional linear aggregators are difficult to capture the complex non-linear relationships between neighbor nodes. Therefore, new GNN models need to be developed that can effectively fuse heterogeneous graph structures and bilinear aggregators to more comprehensively learn the latent feature representations of DDIs.
[0004] (2) How to optimize the GNN model parameters and graph neural network structure to further improve the prediction performance. The performance of GNN models highly depends on their parameters and network structure. Therefore, efficient optimization algorithms need to be developed to automatically learn the optimal parameters and network structure of GNN models to maximize DDI prediction performance. Summary of the Invention
[0005] The technical problem to be solved by the present invention is how to utilize biomedical knowledge and the structure of heterogeneous information network (HIN), and fuse bilinear aggregators to model neighbor nodes, so as to more comprehensively capture the complex semantics of drug-drug interactions (DDIs).
[0006] The present invention realizes the solution to the above technical problem through the following technical means:
[0007] The present invention provides a drug interaction prediction method based on a bilinear graph neural network, including:
[0008] S1. Construct a heterogeneous information network HIN; where different roles of proteins are represented as different types of edges, and drugs and proteins are nodes;
[0009] S2. Establish a bilinear aggregator; perform bilinear combination on the node representations in the heterogeneous information network HIN to generate new nodes;
[0010] S3. Construct an information fusion model MPBGNN based on meta-paths and bilinear aggregators, and define the K-layer MPBGNN as:
[0011]
[0012] where represents the adjacency k-hop connectivity matrix, and GNN K represents a normally recursively defined K-layer GNN. The time complexity of the K-layer MPBGNN is determined by the number of non-zero entries in A (K) ;
[0013] S4. Predict drug interactions; after passing through the l-th layer of the information fusion model MPBGNN based on meta-paths and bilinear aggregators, the final representation matrix of the given drug is A D , and the DDI prediction probability for each pair of drugs is derived by using a fully connected layer with a sigmoid function.
[0014] The present invention utilizes biological knowledge related to the DDI induction process to construct a heterogeneous information network (HIN). In order to capture the complex semantics in the heterogeneous information network, a meta-path-based information fusion mechanism is used to learn high-quality drug representations. In addition, an attention mechanism is designed to combine the semantic information obtained from different-length meta-paths to obtain the final characterization of drugs for DDI prediction.
[0015] As a preferred solution of the above scheme, the different role manifestations of the protein in S1 include:
[0016] (1) The interaction between each pair of proteins, represented by ;
[0017] (2) The protein as a drug target is represented by .
[0018] (3) The interaction between each pair of drugs is represented by .
[0019] (4) The protein acting as the enzyme of the drug is denoted as
[0020] (5) The protein acting as the carrier of the drug is denoted as
[0021] (6) The protein acting as the transporter of the drug is denoted as
[0022] As a preferred solution of the above scheme, the new nodes in S2 are represented as:
[0023]
[0024] where ⊙ is the element product; υ is the target node to obtain the representation; W is the weight matrix for feature transformation; i and j are the node indices of the extended neighbors ; represents the number of interactions of the target node υ, and the obtained representation is normalized to eliminate the bias of the node degree; BA represents the bilinear aggregator, h i is the representation vector h i learned for node i, h j is the representation vector h j learned for node j;
[0025] For the convenience of implementing the matrix calculation, the matrix form of the bilinear aggregator is given as:
[0026]
[0027] As a preferred solution of the above scheme, the specific process of DDI prediction for drugs in S4 is as follows:
[0028] Given drug pairs d i and d j , the final representations a D and a i can be obtained from A j , and then the probability of an interaction between d i and d j is defined as follows:
[0029]
[0030] where [·||·] represents the concatenation of two vectors, and Q and b are the training parameters of the prediction module; in the training phase, assume that the mini-batch input consists of S samples, which are represented as {(X1,y1),…,(X s ,y s ),…,(X S ,y S )}; for the s-th sample, X s =(d is ,d js ), and y s is the true label of X s ; therefore, the cross-entropy loss function is defined as follows:
[0031]
[0032] By minimizing the loss function, all parameters in the MPBGNN model are optimized by the Adam optimizer and in an end-to-end backpropagation manner.
[0033] Corresponding to the above method, the present invention also provides a drug interaction prediction system based on a bilinear graph neural network, including:
[0034] Heterogeneous information network HIN construction module: where different roles of proteins are represented as different types of edges, and drugs and proteins are nodes;
[0035] Bilinear aggregator construction module: bilinearly combines the node representations in the heterogeneous information network HIN to generate new nodes;
[0036] MPBGNN model construction module: constructs an information fusion model MPBGNN based on meta-paths and bilinear aggregators, and defines the K-layer MPBGNN as:
[0037]
[0038] where represents the adjacency k-hop connectivity matrix, and GNN K represents a normally recursively definable K-layer GNN. The time complexity of the K-layer MPBGNN is determined by the number of non-zero entries in A (K) ;
[0039] Drug interaction prediction module: After passing through the l-th layer of the information fusion model MPBGNN based on meta-paths and bilinear aggregators, the final representation matrix of the given drug is A D , and the DDI prediction probability for each pair of drugs is derived by using a fully connected layer with a sigmoid function.
[0040] As a preferred solution of the above scheme, the different role manifestations of proteins in the heterogeneous information network HIN construction module include:
[0041] (1) The interaction between each pair of proteins, represented by ;
[0042] (2) Proteins as drug targets, represented by ;
[0043] (3) The interaction between each pair of drugs, represented by ;
[0044] (4) Proteins acting as enzymes of drugs, denoted as
[0045] (5) Proteins acting as carriers of drugs, denoted as
[0046] (6) Proteins acting as transporters of drugs, denoted as
[0047] As a preferred solution of the above scheme, the new node in the bilinear aggregator construction module is represented as:
[0048]
[0049] where ⊙ is the element-wise product; υ is the target node for which the representation is to be obtained; W is the weight matrix for feature transformation; i and j are the node indices of the extended neighbors ; represents the number of interactions of the target node υ, and the obtained representation is normalized to eliminate the bias of the node degree; BA represents the bilinear aggregator, h i is the representation vector h i learned for node i, h j is the representation vector h j learned for node j;
[0050] For the convenience of implementing matrix calculations, the matrix form of the bilinear aggregator is given as:
[0051]
[0052] As a preferred solution of the above scheme, the specific process of DDI prediction for drugs in the drug interaction prediction module is as follows:
[0053] Given drug pairs d i and d j , the final representations a D and a i can be obtained from A j , and then define d i and dj The probability of interaction between them is as follows:
[0054]
[0055] where [·||·] represents the concatenation of two vectors, and Q and b are the training parameters of the prediction module; in the training stage, assume that the mini-batch input consists of S samples, which are represented as {(X1,y1),…,(X s ,y s ),…,(X S ,y S )}; for the s-th sample, X s =(d is ,d js ), and y s is the true label of X s ; therefore, the cross-entropy loss function is defined as follows:
[0056]
[0057] By minimizing the loss function, all parameters in the MPBGNN model are optimized by the Adam optimizer and in an end-to-end backward propagation manner.
[0058] The present invention also provides a processing device, including at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the above method by calling the program instructions.
[0059] The present invention also provides a computer-readable storage medium, which stores computer instructions, and the computer instructions cause the computer to execute the above method.
[0060] The advantages of the present invention are as follows:
[0061] The present invention utilizes biological knowledge related to the DDI induction process to construct a heterogeneous information network (HIN). To capture the complex semantics in the heterogeneous information network, a meta-path-based information fusion mechanism is used to learn high-quality drug representations. In addition, an attention mechanism is designed to combine the semantic information obtained from different-length meta-paths to obtain the final representation of the drug for DDI prediction.
[0062] The bilinear aggregator of the present invention explicitly encodes the interactions of local nodes to enhance the effect of traditional linear aggregation networks, and is suitable for modeling the neighbor interactions in local structures.
[0063] The present invention uses an effective meta-path-based information fusion mechanism to capture the complex semantic associations between nodes in a heterogeneous information network for learning drug representations. Under the designed attention mechanism, meta-paths of a certain length play a major role in predicting DDIs and provide partial interpretability. Moreover, the present invention incorporates a bilinear aggregation model, which uses pairwise interactions represented by adjacent nodes to increase the weighted sum and explicitly considers the pairwise node interactions in a neat and systematic manner.
[0064] In the constructed heterogeneous information network (HIN), the different roles of proteins are manifested as different types of edges. More specifically, six types of undirected edges are considered, including: (1) the interaction between each pair of proteins; (2) proteins as drug targets; (3) the interaction between each pair of drugs; (4) proteins as enzymes of drugs; (5) proteins as carriers of drugs; (6) proteins as transporters of drugs; By using the same node type (i.e., proteins) to represent proteins with different roles, not only can the redundancy of data storage be reduced, but also the information from different roles of proteins can be conveniently fused.
[0065] The present invention also uses a fully connected layer with a sigmoid function to derive the DDI prediction probability for each pair of drugs, which can handle non-linear decision boundaries and can be easily extended to multi-modal data (such as combined with graph neural networks, text descriptions, etc.) to adapt to more complex DDI prediction scenarios. By minimizing the loss function, all parameters in the proposed model can be optimized by the Adam optimizer and in an end-to-end backward propagation manner. Brief Description of the Drawings
[0066] Figure 1 is a unified framework diagram of a drug interaction prediction method based on a bilinear graph neural network provided by the present invention;
[0067] Figure 2 is an execution flow chart of a drug interaction prediction method based on a bilinear graph neural network. Detailed Embodiments
[0068] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, rather than all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.
[0069] This embodiment describes a drug interaction prediction method based on a bilinear graph neural network, including the following steps:
[0070] Step 1: Construct a heterogeneous information network HIN based on a graph neural network. The constructed DDI-related heterogeneous graph is represented by G=(V, E), where V is the node set and E is the edge set. The node set V can be divided into two subsets S D and S P , representing the drug set and the protein set respectively. Suppose S D ={d1, …, d j , …, d M} and S P ={p1, …, p j , …, p N}, where M and N represent the total number of drugs and proteins respectively. It should be noted that the different roles played by proteins (including targets, carriers, enzymes, and transporters) are not distinguished by node types because the same protein can play multiple roles in different situations. By using the same node type (i.e., protein) to represent proteins with different functions, not only can the redundancy of data storage be reduced, but also the information from different functions of proteins can be conveniently fused.
[0071] Specifically, the construction of the drug-associated heterogeneous graph first requires data collection: collect detailed information on drugs, proteins, and their interactions from drug databases (such as DrugBank) and bioinformatics databases (such as KEGG). Then, perform entity recognition on the collected data to identify entities such as drugs and proteins and map them to the nodes in the HIN. Next, extract various relationships between drugs and proteins from the data, such as DDI, PPI, targets, carriers, enzymes, and transporters, and map them to the edges in the HIN. Finally, construct the identified entities and relationships into a HIN, where drugs and proteins are regarded as heterogeneous types of nodes, and different types of associations (such as DDI, PPI, targets, carriers, enzymes, and transporters) between them are regarded as heterogeneous types of edges.
[0072] In the constructed HIN, the different functions of proteins are manifested as different types of edges. More specifically, consider six types of undirected edges, including:
[0073] (1) The interaction between each pair of proteins, represented by ;
[0074] (2) The protein as a drug target, represented by ;
[0075] (3) The interaction between each pair of drugs, represented by ;
[0076] (4) The protein acting as an enzyme for a drug, denoted as
[0077] (5) The protein acts as a carrier of the drug, denoted as
[0078] (6) The protein acts as a transporter of the drug, denoted as
[0079] Step 2: Establish a bilinear aggregator. Perform a bilinear combination of the node representations in the heterogeneous information network HIN graph to fuse information from different nodes and generate new node representations. The bilinear semantic fusion is specifically the implementation of the bilinear pooling module. First, learn the node features. Using the concept of meta-paths, learn the representations of drugs and proteins. A meta-path describes the composite relationship between node types and can effectively capture the complex semantic associations between nodes in the HIN. Design an information fusion mechanism based on meta-paths to combine the semantic information obtained from different-length meta-paths to obtain the final representation of the drug. Then design an attention mechanism to combine the semantic information obtained from different-length meta-paths to obtain the final representation of the drug for DDI prediction. Then perform graph convolution operations, design a bilinear aggregator, and enhance the weighted sum by calculating the pairwise interactions of adjacent node representations to better capture the interactions between nodes. Use a traditional linear aggregator to perform a weighted sum of the features of neighboring nodes to obtain the representation of the target node. In the output layer, use the learned drug representation to predict the DDI probability between drug pairs through a fully connected layer and a Sigmoid activation function.
[0080] Specifically, first establish the adjacency matrix of the graph to represent the connection relationship between nodes. By multiplying the adjacency matrix with the node feature matrix, obtain the weighted adjacency matrix of each node. Each element represents the weighted contribution of node i to node j, and the weight is determined by the adjacency matrix and the node features.
[0081] Multiply the weighted adjacency matrix by itself to obtain the square of the weighted adjacency matrix. Each element represents the square of the weighted contribution of node i to node j. Continue to multiply the node feature matrix by itself to obtain the sum of the squares of the node features. Multiply the square of the adjacency matrix by the sum of the squares of the node features to obtain the weighted adjacency matrix of the sum of the squares of the node features. Each element represents the weighted contribution of the sum of the squares of the features of node i to node j. Subtract the square of the weighted adjacency matrix from the weighted adjacency matrix of the sum of the squares of the node features and take the average to obtain the new node representation. This new node representation fuses information from different nodes and reflects the complex relationships between nodes.
[0082]
[0083] where ⊙ is the element-wise product; υ is the target node for which the representation is to be obtained; W is the weight matrix (model parameter) for feature transformation; i and j are the extended neighbors The node indices of —— they are constrained to be different to avoid meaningless self-interactions and may even introduce additional noise. Denote the number of interactions of the target node υ, and normalize the obtained representation to eliminate the bias of the node degree.
[0084] For the convenience of implementing matrix calculations, the matrix form of the bilinear aggregator is given as:
[0085]
[0086] where stores the representation vectors h of all nodes, is the adjacency matrix of the graph, where self-loops are added at each node ( is the identity matrix), B is a diagonal matrix, and each element B vv = b v , (·) 2 represents the element-wise product of two matrices.
[0087] Step 3: Construct the information fusion model MPBGNN based on meta-paths and bilinear aggregators. This model takes the l-th layer as an example to describe the idea of meta-path-based information fusion. Without loss of generality, take the drug node d i as an example to explain the node update process in the heterogeneous information network.
[0088] The node d i The representation after the l-th layer of meta-path-based information fusion is defined as follows:
[0089]
[0090] where, and are the training parameters for the l-th layer of information fusion.
[0091] The representation matrix of all drugs after the l-th layer of meta-path information fusion is denoted by . Through the same processing (with different corresponding parameters), the representation matrices of all proteins are also derived and denoted as After the l-th layer of meta-path-based information fusion, the final representation matrices of drugs and proteins are obtained, denoted by H D and H P respectively. Specifically for the first layer, the inputs and are the initial one-hot encodings of drugs and proteins respectively.
[0092] In the implementation of this information network, to reduce the complexity of the proposed model, meta-path-based information fusion is performed according to the length of the meta-path, and at the same time, the node information fused by the bilinear aggregator is added. The maximum length of the considered meta-path is denoted by k. Given the length k ∈ {1, …, k} and the drug node d i , according to all d i with the meta-path length of k, the neighbor nodes can be divided into two subgroups, namely the drug node set (denoted by ) and the protein node set (denoted by ), and ⊙ is the element-wise product. In the meta-path-based information fusion, this mechanism trains different parameters to capture the different contributions of drug neighbors and protein neighbors to the node d i . In the l-th layer, the information fusion of the node d i with the path length of k is formally represented as:
[0093]
[0094] where represents the representation of the node d i under the meta-path with the path length of k after the l-th layer of meta-path-based information fusion, represent the representations of the drug neighbor d t and the protein neighbor p t after the (l - 1)-th layer of meta-path-based information fusion respectively. n d and n p represent the dimensionalities of the representations of drugs and proteins respectively. In addition, ReLU is used as the non-linear activation function, and the parameters and represent the bias vectors of drug neighbors and protein neighbors and the trainable parameter matrix of the l-th layer respectively.
[0095] Spatial Graph Neural Network (GNN) usually learns a representation vector for each node v by recursively aggregating the features of its neighbors to ensure that its features are correctly encoded:
[0096]
[0097] where represents the representation of the target node v at the k-th layer of iteration, W (k) is the weight matrix (model parameter) for feature transformation at the k-th layer, and the initial feature representation can be obtained from the original feature matrix X. The AGG function is usually implemented as a weighted sum, and α viAs the weight of neighbor i. In Graph Convolutional Networks (GCNs), based on Laplace theory, α vi is defined as
[0098] Since the bilinear aggregator emphasizes node interactions and uses a weighted sum aggregator to encode different signals, this model combines them to construct a more expressive graph convolutional network and adopts a simple linear combination scheme to define the new graph convolutional operator as:
[0099] H (k) = (1 - α) · AGG(H (k-1) , A) + α · BA(H (k-1) , A)
[0100] where H (k) stores the node representations of the k-th layer (encoded k-hop neighbors). α is a hyperparameter that balances the advantages of traditional GNN aggregators and bilinear aggregators.
[0101] Since both AGG and BA are permutation-invariant, this graph convolutional operator is also permutation-invariant. When α = 0, node interactions are not considered and MPBGNN degenerates to GNN; when α = 1, MPBGNN only uses the bilinear aggregator to process information from neighbors.
[0102] Traditional GNN models encode information from multi-hop neighbors recursively by stacking multiple aggregators. For example, a two-layer GNN model is formalized as:
[0103]
[0104] where σ is a non-linear activation function. This model designs a two-layer MPBGNN model in the same recursive way:
[0105]
[0106] However, such a straightforward multi-layer extension involves unexpected high-order interactions. In the two-layer case, the second-layer representation will include some fourth-order interactions between 2-hop neighbors, which are difficult to interpret and unreasonable. Therefore, instead of directly stacking MPBGNN layers, the two-layer MPBGNN model is defined as:
[0107] MPBGNN2(X, A) = (1 - α) · GNN2(X, A)
[0108] + α[(1 - β) · BA(X, A) + β · BA(X, A (2) )]
[0109] where A (2) = binarize(AA) stores the 2-hop connectivity of the graph. binarize is an entry operation that converts non-zero entries to 1. Therefore, non-zero entries (v,i) in A (2) mean that node v can reach node i within two hops. β is a hyperparameter used to weigh the strength of the bilinear interaction between 1-hop neighbors and 2-hop neighbors.
[0110] According to the same principle, the K-layer MPBGNN is defined as:
[0111]
[0112] where represents the adjacency k-hop connectivity matrix, and GNN K represents a normally recursively defined K-layer GNN, such as a K-layer GCN or a Graph Attention Network (GAT). The time complexity of the K-layer MPBGNN is determined by the number of non-zero entries in A (K) . To reduce the actual complexity, a sampling strategy in GraphSage can be followed to sample a part of the high-hop neighbors instead of using all neighbors.
[0113] Step 4: Predict drug-drug interactions and optimize related parameters. Finally, DDI prediction and model optimization are performed. According to the trained model that fuses information based on metapaths and bilinear aggregators, the DDI probability between drug pairs is predicted. At the same time, the attention mechanism is used to analyze the contribution of different metapaths to the prediction results, providing partial interpretability for the prediction results. For related optimization, the cross-entropy loss function is used to evaluate the prediction performance of the model, the Adam optimization algorithm is used to optimize the model parameters, and the AUROC and AUPR metrics are used to evaluate the prediction performance of the model.
[0114] Specifically, after passing through the l-th layer of the information fusion model MPBGNN based on metapaths and bilinear aggregators, the final representation matrix of a given drug is A D , and the DDI prediction probability for each pair of drugs is derived by using a fully connected layer with a sigmoid function.
[0115] For a given drug pair d i and d j , the final representations a D and a i can be obtained from A j , and then the probability of an interaction between d i and d j is defined as follows:
[0116]
[0117] where [·||·] represents the concatenation of two vectors, and Q and b are the training parameters of the prediction module. In the training phase, assume that the mini-batch input consists of S samples, which are represented as {(X1,y1),…,(X s ,y s ),…,(X S ,y S )}. For the s-th sample, X s =(d is ,d js ), and y s is the ground-truth label of X s ; d is represents the first feature or dimension of the s-th sample, represents the second feature or dimension of the s-th sample. Therefore, the cross-entropy loss function is defined as follows:
[0118]
[0119] where, is the predicted probability of the s-th sample on the class sum; by minimizing the loss function, all parameters in the proposed model can be optimized by the Adam optimizer and backpropagation in an end-to-end manner.
[0120] The above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements for some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.
Claims
1. A drug interaction prediction method based on a bilinear graph neural network, characterized in that, Including: S1. Construct a heterogeneous information network HIN; wherein different functions of proteins are represented as different types of edges, and drugs and proteins are nodes; S2. Establish a bilinear aggregator; perform a bilinear combination of the node representations in the heterogeneous information network HIN to generate new nodes; S3. Construct an information fusion model MPBGNN based on metapaths and bilinear aggregators, and define the K-layer MPBGNN as: Among them represents the adjacent k-hop connectivity matrix, and GNN K represents a normally recursively definable K-layer GNN. The time complexity of the K-layer MPBGNN is determined by the number of non-zero entries in A (K) ; S4. Predict drug-drug interactions; after passing through the l-th layer of the information fusion model MPBGNN based on meta-path and bilinear aggregator, the final representation matrix of a given drug is obtained as A D , and the DDI prediction probability for each pair of drugs is derived by using a fully connected layer with a sigmoid function.
2. The drug interaction prediction method based on a bilinear graph neural network according to claim 1, wherein The different function manifestations of proteins in S1 include: (1) The interaction between each pair of proteins is represented by ; (2) The protein serving as a drug target is represented by ; (3) The interaction between each pair of drugs is represented by ; (4) The protein acts as the enzyme of the drug, denoted as (5) The protein acts as a drug carrier, denoted as (6) The protein acts as a transporter of the drug, denoted as 3. The drug interaction prediction method based on a bilinear graph neural network according to claim 1, wherein The new node representation in S2 is: where ⊙ is the element product; υ is the target node for which the representation is to be obtained; W is the weight matrix for feature transformation; i and j are the node indices of the extended neighbors of the nodes, denotes the number of interaction times of the target node υ, and the obtained representation is normalized to eliminate the bias of node degree; BA represents a bilinear aggregator, h i is the representation vector h learned for node i i , h j is the representation vector h learned for node j j ; For the convenience of matrix calculation implementation, the matrix form of the bilinear aggregator is given as: Among them The representation vectors h of all nodes are stored, is the adjacency matrix of the graph, where self-loops are added to each node is the identity matrix, B is a diagonal matrix, and each element B vv = b v , (·) 2 represents the element-wise product of two matrices.
4. The method for predicting drug interactions based on a bilinear graph neural network according to any one of claims 1 to 3, characterized in that The specific process of DDI prediction for drugs in S4 is: For a given drug pair d i and d j ,a final representation a D can be obtained from A i and a j ,and then the probability of interaction between d i and d j is defined as follows: where [·||·] represents the concatenation of two vectors, and Q and b are the training parameters of the prediction module; during the training phase, assume that the mini-batch input consists of S samples, which are represented as {(X1,y1),…,(X s ,y s ),…,(X S ,y S )}; for the s-th sample, y s is the true label of X s ; therefore, the cross-entropy loss function is defined as follows: By minimizing the loss function, all parameters in the MPBGNN model are optimized through the Adam optimizer and in an end-to-end backpropagation manner.
5. A drug interaction prediction system based on a bilinear graph neural network, characterized in that, Including: Heterogeneous information network HIN construction module: wherein different functions of proteins are represented as different types of edges, and drugs and proteins are nodes; Bilinear aggregator construction module: perform a bilinear combination of the node representations in the heterogeneous information network HIN to generate new nodes; MPBGNN model construction module: construct an information fusion model MPBGNN based on metapaths and bilinear aggregators, and define the K-layer MPBGNN as: Among them represents the adjacent k-hop connectivity matrix, and GNN K represents a normally recursively definable K-layer GNN. The time complexity of the K-layer MPBGNN is determined by the number of non-zero entries in A (K) ; Drug interaction prediction module: After passing through the l-th layer of the information fusion model MPBGNN based on meta-path and bilinear aggregator, the final representation matrix of the given drug is obtained as A D , and the DDI prediction probability of each pair of drugs is derived by using a fully connected layer with a sigmoid function 6. The drug interaction prediction system based on the bilinear graph neural network according to claim 5, characterized in that, The different function manifestations of proteins in the heterogeneous information network HIN construction module include: (1) The interaction between each pair of proteins is represented by ; (2) The protein as a drug target is represented by ; (3) The interaction between each pair of drugs is represented by ; (4) The protein acts as the enzyme of the drug, denoted as (5) The protein acts as a carrier for the drug, denoted as (6) The protein acts as a transporter of the drug, denoted as 7. The drug interaction prediction system based on the bilinear graph neural network according to claim 5, characterized in that, The new node representation in the bilinear aggregator construction module is: where ⊙ is the element product; υ is the target node for which the representation is to be obtained; W is the weight matrix for feature transformation; i and j are the node indices of the extended neighbors of the node indices representing the number of interaction times of the target node υ, and the obtained representation is normalized to eliminate the bias of the node degree; BA represents a bilinear aggregator, h i is the representation vector h learned for node i i , h j is the representation vector h learned for node j j ; For the convenience of matrix calculation implementation, the matrix form of the bilinear aggregator is given as:
8. The drug interaction prediction system based on the bilinear graph neural network according to any one of claims 5 to 7, characterized in that The specific process of DDI prediction for drugs in the drug interaction prediction module is: For a given drug pair d i and d j it is possible to obtain the final representations a D and a i from A j and then define the probability of interaction between d i and d j as follows: where [·||·] represents the concatenation of two vectors, and Q and b are the training parameters of the prediction module; in the training phase, assume that the mini-batch input consists of S samples, which are represented as {(X1,y1),…,(X s ,y s ),…,(X S ,y S )}; for the s-th sample, y s is the true label of X s ; therefore, the cross-entropy loss function is defined as follows: By minimizing the loss function, all parameters in the MPBGNN model are optimized through the Adam optimizer and in an end-to-end backpropagation manner.
9. A processing device, characterized in that, Including at least one processor and at least one memory communicatively connected to the processor, wherein: the memory stores program instructions executable by the processor, and the processor can execute the method according to any one of claims 1 to 4 by invoking the program instructions.
10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions, and the computer instructions cause the computer to execute the method according to any one of claims 1 to 4.