Multi-view drug pair reaction prediction method, device, equipment and storage medium

Abstract

The application provides a multi-view drug pair reaction prediction method, device, equipment and storage medium, and relates to the field of drug pair reaction prediction. The drug pair reaction prediction method comprises the following steps: S1, obtaining feature information and known reaction information of a to-be-predicted drug; S2, constructing an adjacency view according to the known reaction information; S3, constructing a network according to the known reaction information, and constructing a diffusion view; S4, constructing a drug feature matrix and a KNN view according to the feature information; S5, performing deep fusion of the drug feature matrix and the adjacency view, the diffusion view and the KNN view respectively through a node-level attention network to obtain three view-level embedding vectors; S6, fusing the three view-level embedding vectors to obtain embedding representations of the drugs; S7, combining the embedding representations of the drugs to obtain a plurality of drug pair vectors; S8, taking drug pair vectors of a known reaction type as a training set to train a classification model; and S9, inputting drug pair vectors of an unknown reaction type into the classification model to obtain a predicted reaction type.

Description

Technical Field

[0001] This invention relates to the field of drug pair response prediction technology, and more specifically, to a multi-view drug pair response prediction method, apparatus, device, and storage medium. Background Technology

[0002] Combination therapy with multiple drugs is a common approach in clinical practice and has proven to be a safe and effective method for treating serious diseases. However, the use of multiple drugs may also increase the risk of adverse events, posing safety risks to patients. Therefore, the prediction of drug interactions (DDIs) plays a crucial role in pharmacological research and clinical applications, making the early identification of DDIs an urgent task.

[0003] Currently, most methods for identifying DDI on the market rely on in vivo and in vitro clinical trials, which are time-consuming and cumbersome. With the advancement of artificial intelligence, computational methods for predicting DDI have seen widespread development. These methods can be broadly categorized as follows: (1) feature similarity-based methods; (2) graph embedding-based methods; and (3) attribute network methods. The first method is natural and direct, but it ignores structural information. Furthermore, the selection of features depends on expert experience. The second method ignores feature information, which can provide valuable discriminative information for the prediction task. The third method is a shallow model that cannot capture the nonlinear relationship between features and topological information, and it does not effectively combine features with topological information. Because shallowness is only one type of method, it is merely a manifestation and some interpretations are biased.

[0004] Furthermore, prior art DDI prediction methods generally exhibit weak robustness. Therefore, the applicant hereby submits this application after reviewing existing technologies. Summary of the Invention

[0005] The present invention provides a multi-view drug pair response prediction method, apparatus, device, and storage medium to improve at least one of the above-mentioned technical problems.

[0006] First aspect

[0007] This invention provides a multi-view drug pair response prediction method, which includes steps S1 to S9.

[0008] S1. Obtain the characteristic information of the drug to be predicted and the known reaction information between drugs;

[0009] S2. Construct an adjacency view based on the known reaction information;

[0010] S3. Construct a network based on the known reaction information, and use a personalized random walk PPR to construct a diffusion view;

[0011] S4. Construct a drug feature matrix based on the feature information; and construct a view based on the drug feature matrix to obtain a KNN view;

[0012] S5. Through a node-level attention mechanism, the drug feature matrix is ​​deeply fused with the adjacency view, the diffusion view, and the KNN view to obtain three view-level embedding vectors.

[0013] S6. By fusing the three view-level embedding vectors through a view-level attention network, the embedding representation of each drug is obtained.

[0014] S7. Combine the embedded representations of each drug to obtain multiple drug pair vectors; wherein the response type of some drug pair vectors is known.

[0015] S8. Train the classification model using drug pair vectors with known reaction types as the training set;

[0016] S9. Input the drug pair vector with unknown reaction type into the classification model to obtain the predicted reaction type.

[0017] The second aspect

[0018] This invention provides a multi-view drug pair response prediction device, comprising:

[0019] The initial information acquisition module is used to acquire the characteristic information of the drug to be predicted and the known reaction information between drugs;

[0020] The first view acquisition module is used to construct an adjacency view based on the known reaction information;

[0021] The second view acquisition module is used to construct a network based on the known reaction information and to construct a diffusion view using a personalized random walk PPR.

[0022] The third view acquisition module is used to construct a drug feature matrix based on the feature information; and to construct a view based on the drug feature matrix to obtain a KNN view;

[0023] The view-level vector acquisition module is used to perform deep fusion of the drug feature matrix with the adjacent view, the diffuse view, and the KNN view through a node-level attention mechanism to obtain three view-level embedding vectors.

[0024] The fusion module is used to fuse the three view-level embedding vectors through a view-level attention network to obtain the embedding representation of each drug.

[0025] The combination module is used to combine the embedded representations of the various drugs to obtain multiple drug pair vectors; wherein the reaction types of some drug pair vectors are known.

[0026] The training module is used to train a classification model using drug pair vectors with known response types as the training set.

[0027] The prediction module is used to input drug pair vectors with unknown reaction types into the classification model to obtain the predicted reaction types.

[0028] Third aspect

[0029] This invention provides a multi-view drug pair response prediction device, which includes a processor, a memory, and a computer program stored in the memory; the computer program can be executed by the processor to implement the multi-view drug pair response prediction method as described in any paragraph of the first aspect.

[0030] Fourth aspect

[0031] This invention provides a computer-readable storage medium. The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the multi-view drug pair response prediction method as described in any paragraph of the first aspect.

[0032] By adopting the above technical solution, the present invention can achieve the following technical effects:

[0033] The drug pair response prediction method of the present invention improves accuracy, scalability and robustness by constructing multiple views to capture comprehensive topological information. Attached Figure Description

[0034] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is a flowchart illustrating a method for predicting drug response.

[0036] Figure 2 This is a logical diagram of a drug response prediction method.

[0037] Figure 3 This is a schematic diagram of a drug response prediction device. Detailed Implementation

[0038] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0039] Example 1

[0040] Please see Figure 1 and Figure 2 The first embodiment of the present invention provides a multi-view drug pair response prediction method, which can be executed by a drug pair response prediction device (hereinafter referred to as: prediction device). In particular, it is executed by one or more processors in the prediction device to implement steps S1 to S9.

[0041] S1. Obtain the characteristic information of the drug to be predicted and the known reaction information between drugs.

[0042] Specifically, the aforementioned feature information and known response information can be obtained through methods such as internet searches, and this invention does not limit the specific method of acquisition. It is understood that the prediction device can be an electronic device with computing power, such as a portable laptop, desktop computer, server, smartphone, or tablet computer.

[0043] Based on the above embodiments, in an optional embodiment of the present invention, step S1 specifically includes: acquiring the feature information of the drug to be predicted and the known reaction information between drugs. Then, based on the feature information and the known reaction information, a multi-relationship network graph G is constructed. Wherein, G = {V, E, X}, where V represents drug nodes, E represents the reactions between drugs, and X represents the features of the drugs.

[0044] Preferably, the experimental data used in this embodiment is derived from data contributed by the paper "A multimodal deep learning framework for predicting drug-drug interaction events. Bioinformatics" (hereinafter referred to as DDIMDL). This paper provides four characteristics of 572 drugs and 37,264 drug-drug interaction events. The four characteristics include chemical substructures, targets, enzymes, and pathways. The reaction events involve 65 types. Based on the experimental conclusions of DDIMDL, this embodiment selects the three characteristics with the best results: chemical substructures, targets, and enzymes.

[0045] This embodiment uses the relevant symbols in Table 1 to define the problem.

[0046] Table 1. Definitions of Main Symbols

[0047]

[0048] Specifically, based on the obtained characteristic information of the drug to be predicted and the known reaction information between drugs, a multi-relationship DDI network diagram is constructed.

[0049] G = {V, E, X}

[0050] In the formula, V represents the drug node, E represents the reaction between drugs, and X represents the characteristics of the drug.

[0051] The task in this embodiment is to predict the response type between two drugs and a drug (dj), which is essentially defined as a multi-class classification problem. The output of the result represents the probability distribution of the events predicted by the drug (DDI), and the category with the highest score is the final response type.

[0052] S2. Construct an adjacency view based on the known reaction information.

[0053] Specifically, based on the multi-relationship DDI network graph G, different structures are extracted from the drug relationship network (i.e., different topological structures are extracted from different perspectives / aspects). These different views can represent different aspects of the relationship between drugs.

[0054] Based on the above embodiments, in an optional embodiment of the present invention, step S2 specifically includes: constructing an adjacency view A according to the multi-relationship network graph G. adj In the adjacent view, if two drugs are known to interact, then in A... adj Set the corresponding position in the matrix to 1, otherwise set it to 0.

[0055] Specifically, adjacent view A adj It reflects local topological information and is constructed from real reaction relationships. If two drugs interact, then A adj The corresponding position in the matrix will be replaced with "1"; otherwise, it will be replaced with "0".

[0056] S3. Construct a network based on the known reaction information, and use a personalized random walk PPR to construct a diffusion view.

[0057] Based on the above embodiments, in an optional embodiment of the present invention, step S3 specifically includes: constructing a diffusion view A based on the multi-relationship network graph G. diff Among them, diffusion view A diff The construction model is as follows: In the formula, α∈(0,1] represents the instantaneous transition probability in the random walk, I is the identity matrix, A is the adjacency matrix, and D is the degree matrix of the adjacency matrix A.

[0058] Specifically, diffuse view A diff A global view of drug-reaction relationships. It can obtain topological information beyond first-order neighbors to predict DDI. Diffusion View A diff The formula for calculating Personal PageRank (PPR) is:

[0059]

[0060] In the formula, ∈(0,1] represents the instantaneous transition probability in the random walk, I is the identity matrix, A is the adjacency matrix, and D is the degree matrix of the adjacency matrix A.

[0061] S4. Construct a drug feature matrix based on the feature information; and construct a KNN view based on the drug feature matrix.

[0062] Specifically, the characteristic information of a drug is its most important information, and it needs to be encoded and converted into a vector format that can be recognized and calculated by a computer. Based on the above embodiments, in an optional embodiment of the present invention, step S4 specifically includes steps S41 to S45.

[0063] S41. Based on the feature information, construct the initial feature matrix for each feature type.

[0064] Specifically, each drug can be represented by a binary vector, and each feature corresponds to a set of feature descriptors, whose value (1 or 0) indicates the presence or absence of the corresponding descriptor. Taking chemical structures as an example: for instance, PubChem defines 583 seed structures, so an initial feature matrix of chemical structures with the shape (572, 583) can be obtained; where 572 is the number of drugs and 583 is the number of chemical structures.

[0065] By applying the same treatment to other features (targets, enzymes), the initial feature matrix of the other features can be obtained.

[0066] S42. Perform Jaccard similarity calculations on each initial feature matrix to obtain the similarity matrix for each feature type.

[0067] Specifically, the feature vectors of drugs in the initial feature matrix have high dimensionality and many "0" values, which may lead to reduced model performance. Therefore, in this embodiment, a similarity matrix is ​​obtained through Jaccard similarity calculation to alleviate this issue. The Jaccard similarity calculation model is as follows:

[0068]

[0069] In the formula, di and dj are binary vectors of drugs i and j, di∩dj represents the intersection, and di∪dj represents the union.

[0070] The same process was applied to other features, resulting in three similarity matrices X, each of size (572, 572). s ,X t ,X e .

[0071] S43. Concatenate the similarity matrices of each feature type to obtain the drug feature matrix. The drug feature matrix contains the feature vectors of each drug to be predicted.

[0072] Specifically, the similarity matrices corresponding to each feature type are concatenated to obtain the final vector representation X of the drug. The final vector representation contains the feature vectors of each drug.

[0073]

[0074] In the formula, X di Let i represent the feature vector of drug i, which has 1716 dimensions. The symbol ⊕ indicates a concatenation operation.

[0075] S44. Based on the drug feature matrix, calculate the cosine similarity between nodes, and select the K nodes with the highest similarity as the nodes' neighbors. The cosine similarity calculation model is as follows:

[0076]

[0077] In the formula, s ij Represents the similarity between vectors i and j, X di The feature vector of drug i, X dj Let || denote the feature vector of drug j, and || denote the Euclidean normal operation.

[0078] In this embodiment, K is set to 10. In other embodiments, K can be set to other values ​​as needed, and the present invention does not impose specific limitations on this.

[0079] S45. Connect an edge between the node and its neighbor to obtain the KNN view A. knn .

[0080] Specifically, KNN diagram A knn This reflects the similarity of the feature space. In this embodiment, the cosine similarity between node i and other nodes is calculated, and then the top K nodes with the highest cosine similarity are selected as the neighbor nodes of node i. Then, an edge is connected between each node and its top ten neighbors, thus constructing the KNN graph.

[0081] S5. Through a node-level attention mechanism (GAT model), the drug feature matrix is ​​deeply fused with the adjacency view, the diffusion view, and the KNN view to obtain three view-level embedding vectors.

[0082] Specifically, fusing feature information and topological relationship information can greatly enrich the information contained in the embedded vector. Based on the above embodiments, in an optional embodiment of the present invention, step S5 specifically includes steps S51 to S54.

[0083] S51. Combine the drug feature matrix with the adjacency view, diffusion view, and KNN view respectively. Figure 1 The input cross-fusion network is used to perform deep fusion by performing the following steps to obtain three view-level embedding vectors:

[0084] S52, the cross-fusion network extracts feature vectors at different levels from the drug feature matrix through an autoencoder network.

[0085] S53. The cross-fusion network extracts view vectors from different levels of the view through a graph attention network. Specifically, the feature vectors extracted from each layer of the autoencoder network are transmitted to the corresponding layers of the graph attention network for fusion. Preferably, step S53 specifically includes: obtaining the node attention coefficients of neighboring nodes through a self-attention mechanism, and updating the node representation based on the node attention coefficients. The node representation update model is as follows: In the formula, h i Let be the feature representation of the updated node i, σ be the activation function, and N be the value of N. i Let W represent the set of neighbors of node i, W be the parameter matrix, and h be the number of neighbors of node i. j The feature representation of neighbor node j.

[0086] Specifically, the importance of different neighboring nodes varies. This embodiment adaptively learns the weights using a self-attention mechanism. Taking node i as an example: the influence of neighbor j on i is calculated using the following formula:

[0087]

[0088] In the formula, N represents the neighbors of node i, and the attention coefficient α ij It is e ij The result is obtained after normalization using softmax. ij The following formula is used to calculate:

[0089] e ij =LeakyReLU(w T [Wh i ||Wh j ])

[0090] In the formula, the || symbol means concatenation, T represents transpose, W is the parameter matrix, and h i h j Let i and j represent the characteristics of drugs i and j.

[0091] Similarly, we can obtain the attention scores of other neighboring nodes for drug i. Finally, we update the node representation using a weighted summation, formally expressed as follows:

[0092]

[0093] S54. By fusing the feature vector extracted from the last layer of the autoencoder network and the view vector extracted from the last layer of the graph attention network through the attention mechanism, a view-level embedding vector is obtained.

[0094] Specifically, to efficiently integrate network structure and node attribute information, a cross-fusion strategy is employed for each view. For clarity, the following example uses A... adj Let's take a view as an example:

[0095] like Figure 2 As shown, given the feature matrix X and the topological information matrix A adj Two types of information are encoded using a dual-channel approach.

[0096] For feature information, an autoencoder network consisting of multiple sequentially connected autoencoder models (AEs) is selected to learn the latent embeddings. The feature embeddings learned in the l-th layer are represented by H. (l) express:

[0097]

[0098] In the formula, Table ReLU activation function, W (l) and b (l) These are the weight matrix and bias matrix of the l-th layer of the drug feature vector in the encoder, respectively. Where H... (0) Let X represent the initial feature matrix after combination.

[0099] For capturing topological information, we consider the varying importance of different neighbors to the central node. During aggregation, we employ a node-level attention mechanism to assign different weights to neighbors, implemented using a graph attention network composed of multiple sequentially connected graph attention models (GATs). The topological embedding learned at each layer is then processed using... express.

[0100] This embodiment sets up a transmission operation to flexibly exchange and fuse information between the two channels. Specifically, features extracted from the AE are embedded and transmitted to the GAT network with a certain probability α, formally expressed as follows:

[0101]

[0102] In the formula, α is the fusion coefficient, which is used to balance the importance of vectors learned from the GAT model and the AE model.

[0103] Then, This is then treated as a new feature and input into the next GAT layer to obtain a fused vector representation:

[0104]

[0105] In the formula, Table of ReLU activation functions, The GAT channel is the fused embedding representation of the l-th layer.

[0106] Finally, to maximize the use of information from both channels, the two embedding vectors are fused using an attention mechanism before the output layer. The fusion process using the attention mechanism is similar to step S6.

[0107] The drug feature matrix is ​​then compared with the adjacency view, the diffusion view, and the KNN view. Figure 1 The input is fed into a cross-fusion network, and then the above steps are used to obtain three view-level embedding vectors Z. adj Z diff Z diff These three embeddings are rich in information because they simultaneously contain node features and topological relationship information from the first l layers. This high-quality embedding table representation lays a solid foundation for downstream prediction tasks.

[0108] S6. By using a view-level attention network, fuse three view-level embedding vectors to obtain the embedding representation of each drug.

[0109] Preferably, step S6 specifically includes: obtaining the view attention coefficients of the updated view-level embedding vectors, and fusing the three updated view-level embedding vectors according to the view attention coefficients to obtain the embedding representation of each drug.

[0110] Specifically, this embodiment employs view-level attention, assigning low weights to view embeddings with low contribution and high weights to important views. This helps reduce noise and achieve more efficient fusion.

[0111] First, the attention coefficients of the three graph embeddings are calculated. In this embodiment, ε is used. a ,ε k and ε d The subscript indicates the first letter of the corresponding view, as shown in the following formula:

[0112] (ε a ,ε d,ε k ) = att(Z adj Z diff Z knn )

[0113] With ε a Taking the calculation process as an example: Drug i is embedded in Z in the adjacency graph adj The representation in is First, perform a non-linear transformation on it, then multiply it with the shared attention vector q to obtain its attention value. As shown in the following formula:

[0114]

[0115] In the formula, w is the weight matrix and b is the bias vector.

[0116] Similarly, we can obtain the attention values ​​of drug i embedded in the other two views.

[0117] Then, the attention value is regularized using Softmax to obtain its final weight coefficients, as shown in the following formula:

[0118]

[0119] Similarly, we can obtain the attention scores for the other two views. and For all drugs, the attention scores are collected, and the weighting coefficients can be expressed as ε. t =[ε t ], ε k =[ε k ], ε d =[ε d ], after diagonalization, can be represented as ε t =diag(ε t ), ε k =diag(ε k ), ε d =diag(ε d ).

[0120] Finally, the unified drug embedding representation vector Z is obtained as shown in the following equation:

[0121] Z = ε a ·Z adj +ε d ·Z diff +ε k ·Z knn

[0122] S7. Combine the embedding representations of each drug to obtain multiple drug pair vectors. Among them, the response type of some drug pair vectors is known.

[0123] Specifically, the deep fusion in steps S4 and S5 has yielded the embedded representation Z of all drugs. The vectors of two drugs i and j are represented by Ф(di) and Ф(dj), respectively. This embodiment constructs drug pair vector representations using four different methods, as shown in Table 2.

[0124] Table 2. Four different drug pair combinations

[0125]

[0126] Where the ⊙ symbol represents the hadamard operation, the symbol Table concatenation joins. The dimensions only change when the concatenation join method is selected.

[0127] It is understood that other combinations may be used in other embodiments, and are not limited to the four combinations described above.

[0128] S8. Use drug pair vectors with known reaction types as the training set to train the classification model.

[0129] Specifically, in the preceding steps, the amount of drug has been extracted. To predict the response type of the drug pair, a corresponding classification model needs to be trained. The training dataset consists of drug pair vectors representing the response type. Preferably, based on the above embodiments, in an optional embodiment of the present invention, S8 specifically includes steps S81 to S83.

[0130] S81. Obtain all types of interactions between drug pairs.

[0131] S82. Obtain the drug pair vector with known reaction types.

[0132] S83. Using the reaction type between drug pairs as labels, the drug pair vectors with known reaction types as the training dataset, and reconstruction loss and cross-entropy loss functions as loss functions, the model is trained to obtain the classification model.

[0133] Specifically, in this embodiment, the cross-training loss is defined as l. ce and reconstruction loss l re Optimize the model.

[0134] For node attributes, the decoder reconstructs the data directly from the learned embeddings, and the reconstruction loss is defined as follows:

[0135]

[0136] In the formula, N represents the number of drugs, and xi and These represent the original attribute features and reconstructed features of node i, respectively.

[0137] For multi-class classification tasks, we use the cross-entropy loss function for evaluation. The cross-entropy loss function is:

[0138]

[0139] In the formula, L represents the set of drug pair vectors in the training set, and c represents the category of drug response events. With Y∈R L×C Let these be the predicted labels and the set of actual labels for each drug pair, respectively.

[0140] It should be noted that the classification model can be any classification model in the prior art, and this invention does not specifically limit it.

[0141] S9. Input the drug pair vector with unknown reaction type into the classification model to obtain the predicted reaction type.

[0142] The drug-pair response prediction method (MDFA model architecture, such as) of this invention embodiment Figure 2 As shown in the figure, it mainly consists of four parts: multi-view construction, deep fusion, two-layer attention fusion, and model optimization.

[0143] It first constructs three views A from different angles. adj A diff A knn Under each view, a specific view A y The feature information X is input into the corresponding network. To efficiently integrate the two types of information, this paper adopts a cross-fusion strategy. Specifically, the graph attention network (GAT) model is applied to the relation matrix to capture topological information, and the autoencoder (AE) model is applied to the feature matrix to capture feature information. Then, the two types of information are cross-fused through a transmission operation at each layer. After multiple layers of transmission, the embedding representation Z of the graph below is obtained in three different views. adj Z diff Z knn They are then fused through an attention mechanism to obtain a unified drug embedding representation. Finally, this is constructed into multiple drug pair vectors, which are input into a loss function for training and optimization.

[0144] Through extensive research, the inventors discovered that prior art primarily focuses on a single view of drug-drug interactions, neglecting the importance of multiple views. Therefore, the model's results are highly dependent on the quality of the selected single view, leading to weak robustness.

[0145] The drug-pair response prediction method of this invention constructs multiple views of drug interaction relationships, leveraging the complementarity of these views to capture more comprehensive and accurate topological information, thus improving accuracy, scalability, and robustness. Furthermore, it achieves effective fusion of drug nodes with different views by fusing embedding vectors at three view levels through a two-layer attention mechanism. The attention score of the two-layer attention mechanism also provides some interpretability for drug responses. This drug-pair response prediction method can predict multiple drug response events, which is more conducive to studying the mechanisms hidden behind drug responses.

[0146] The drug-pair response prediction method of this invention integrates information from multiple views, reducing reliance on a single view and significantly improving model stability. Furthermore, within each view, transitive operations are used to flexibly exchange feature information with topological information, fully capturing interaction information and obtaining high-quality drug characterization. Additionally, a bi-level attention mechanism can also provide interpretable predictions of drug responses.

[0147] Example 2

[0148] like Figure 3 As shown, an embodiment of the present invention provides a multi-view drug pair response prediction device, which includes:

[0149] Initial information acquisition module 1 is used to acquire the characteristic information of the drug to be predicted and the known reaction information between drugs.

[0150] First view acquisition module 2 is used to construct an adjacency view based on known reaction information.

[0151] The second view acquisition module 3 constructs a network based on known reaction information and uses a personalized random walk PPR to construct a diffusion view.

[0152] The third view acquisition module 4 is used to construct a drug feature matrix based on feature information, and then construct a view based on the drug feature matrix to obtain a KNN view.

[0153] The view-level vector acquisition module 5 is used to deeply fuse the drug feature matrix with the adjacent view, the diffuse view, and the KNN view through a node-level attention mechanism to obtain three view-level embedding vectors.

[0154] Fusion module 6 is used to fuse three view-level embedding vectors through a view-level attention network to obtain the embedding representation of each drug.

[0155] Combination module 7 is used to combine the embedded representations of various drugs to obtain multiple drug pair vectors. Among them, the response types of some drug pair vectors are known.

[0156] Training module 8 is used to train a classification model using drug pair vectors with known reaction types as the training set.

[0157] Prediction module 9 is used to input drug pair vectors with unknown reaction types into the classification model to obtain the predicted reaction type.

[0158] Based on the above embodiments, in an optional embodiment of the present invention, the view-level vector acquisition module 4 includes:

[0159] Combination units are used to combine the drug feature matrix with the adjacency view, the diffusion view, and the KNN view, respectively. Figure 1 The input cross-fusion network is used to perform deep fusion by performing the following steps to obtain three view-level embedding vectors:

[0160] The feature extraction unit is used by the cross-fusion network to extract feature vectors at different levels from the drug feature matrix through the autoencoder network.

[0161] The structure extraction unit is used to extract view vectors at different levels from the view through the graph attention network in the cross-fusion network. Specifically, the feature vectors extracted by each layer of the autoencoder network are transmitted to the corresponding layers of the graph attention network for fusion.

[0162] The fusion unit is used to fuse the feature vector extracted from the last layer of the autoencoder network and the view vector extracted from the last layer of the graph attention network through an attention mechanism to obtain a view-level embedding vector.

[0163] Based on the above embodiments, in an optional embodiment of the present invention, the structure extraction unit is configured to: obtain the node attention coefficients of neighboring nodes through a self-attention mechanism, and update the node representation according to the node attention coefficients; wherein, the node representation update model is: In the formula, h i Let be the feature representation of the updated node i, σ be the activation function, and N be the value of N. i Let W represent the set of neighbors of node i, W be the parameter matrix, and h be the number of neighbors of node i. j The feature representation of neighbor node j.

[0164] Based on the above embodiments, in an optional embodiment of the present invention, the initial information acquisition module 1 is specifically used to: acquire the feature information of the drug to be predicted and the known reaction information between drugs. And based on the feature information and the known reaction information, a multi-relationship network graph G is constructed. Wherein, G = {V, E, X}, where V represents drug nodes, E represents the reactions between drugs, and X is the feature of the drug.

[0165] Based on the above embodiments, in an optional embodiment of the present invention, the first view acquisition module 2 is specifically used to: construct an adjacency view A based on the multi-relationship network graph G. adj In the adjacent view, if two drugs are known to interact, then in A... adj Set the corresponding position in the matrix to 1, otherwise set it to 0.

[0166] Based on the above embodiments, in an optional embodiment of the present invention, the second view acquisition module 3 is specifically used to: construct a diffusion view A based on the multi-relationship network graph G. diff Among them, diffusion view A diff The construction model is as follows: In the formula, α∈(0,1] represents the instantaneous transition probability in the random walk, I is the identity matrix, A is the adjacency matrix, and D is the degree matrix of the adjacency matrix A.

[0167] Based on the above embodiments, in an optional embodiment of the present invention, the third view acquisition module 4 specifically includes:

[0168] The initial feature matrix construction unit is used to construct the initial feature matrix for each feature type based on the feature information.

[0169] The similarity matrix calculation unit is used to perform Jaccard similarity calculations on each initial feature matrix to obtain the similarity matrix for each feature type.

[0170] The concatenation unit is used to concatenate the similarity matrices of various feature types to obtain the drug feature matrix. The drug feature matrix contains the feature vectors of each drug to be predicted.

[0171] The similarity calculation unit is used to calculate the cosine similarity between nodes based on the drug feature matrix, and to obtain the neighbors of each node. The calculation model for cosine similarity is as follows: In the formula, s ij Represents the similarity between vectors i and j, X di The feature vector of drug i, X dj Let || denote the feature vector of drug j, and || denote the Euclidean normal operation.

[0172] The third-view construction unit is used to connect an edge between a node and its neighbors to obtain the KNN view A. knn .

[0173] Based on the above embodiments, in an optional embodiment of the present invention, the fusion module 6 is specifically used to: obtain the view attention coefficients of the updated view-level embedding vectors, and fuse the three updated view-level embedding vectors according to the view attention coefficients to obtain the embedding representation of each drug.

[0174] Example 3

[0175] This invention provides a multi-view drug pair response prediction device, which includes a processor, a memory, and a computer program stored in the memory. The computer program can be executed by the processor to implement the multi-view drug pair response prediction method as described in any paragraph of Embodiment 1.

[0176] Example 4

[0177] This invention provides a computer-readable storage medium. The computer-readable storage medium includes a stored computer program, wherein, when the computer program is executed, it controls the device containing the computer-readable storage medium to perform the multi-view drug pair response prediction method as described in any paragraph of Embodiment 1.

[0178] In the several embodiments provided in this invention, it should be understood that the disclosed apparatus and methods can also be implemented in other ways. The apparatus and method embodiments described above are merely illustrative. For example, the flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of apparatus, methods, and computer program products according to various embodiments of the present invention. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions marked in the blocks may occur in a different order than those marked in the drawings. For example, two consecutive blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in a block diagram and / or flowchart, and combinations of blocks in block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or action, or using a combination of dedicated hardware and computer instructions.

[0179] In addition, the functional modules in the various embodiments of the present invention can be integrated together to form an independent part, or each module can exist independently, or two or more modules can be integrated to form an independent part.

[0180] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, electronic device, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks. It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. In the absence of further restrictions, an element defined by the phrase "comprising a..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0181] The terminology used in the embodiments of this invention is for the purpose of describing particular embodiments only and is not intended to limit the invention. The singular forms “a,” “the,” and “the” as used in the embodiments of this invention and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0182] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0183] Depending on the context, the word "if" as used here can be interpreted as "when," "when," "in response to determination," or "in response to detection." Similarly, depending on the context, the phrase "if determination" or "if detection (of the stated condition or event)" can be interpreted as "when determination," "in response to determination," "when detection (of the stated condition or event)," or "in response to detection (of the stated condition or event)."

[0184] The use of "first" and "second" in the embodiments is merely to distinguish similar objects and does not represent a specific ordering of objects. It is understood that "first" and "second" can be interchanged in a specific order or sequence where permitted. It should be understood that the objects distinguished by "first" and "second" can be interchanged where appropriate so that the embodiments described herein can be implemented in an order other than those illustrated or described herein.

[0185] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-view drug pair reaction prediction method, characterized in that, The method comprises the following steps: obtaining feature information of a drug to be predicted and known reaction information between drugs; constructing an adjacency view according to the known reaction information; constructing a network according to the known reaction information, and constructing a diffusion view by using personalized random walk PPR; constructing a drug feature matrix according to the feature information; and constructing a KNN view according to the drug feature matrix; by a node-level attention network, the drug feature matrix is respectively and deeply fused with the adjacency view, the diffusion view, and the KNN view to obtain three view-level embedding vectors; by a view-level attention network, the three view-level embedding vectors are fused to obtain an embedding representation of each drug; combining the embedding representations of the drugs to obtain a plurality of drug pair vectors; wherein the reaction type of part of the drug pair vectors is known; using the drug pair vectors with known reaction types as a training set to train a classification model; inputting the drug pair vectors with unknown reaction types into the classification model to obtain a predicted reaction type; by a node-level attention network, the drug feature matrix is respectively and deeply fused with the adjacency view, the diffusion view, and the KNN view to obtain three view-level embedding vectors, specifically comprising: the drug feature matrix is input into a cross-fusion network together with the adjacency view, the diffusion view, and the KNN view to perform the following steps for deep fusion to obtain three view-level embedding vectors: the cross-fusion network extracts different levels of feature vectors from the drug feature matrix through a self-encoder network; the cross-fusion network extracts different levels of view vectors from the views through a graph attention network; wherein the feature vectors extracted by each layer of the self-encoder network are transmitted to the corresponding layer of the graph attention network for fusion; by an attention mechanism, the feature vector extracted by the last layer of the self-encoder network and the view vector extracted by the last layer of the graph attention network are fused to obtain a view-level embedding vector.

2. The multi-view drug pair reaction prediction method of claim 1, wherein, the cross-fusion network extracts different levels of view vectors from the views through a graph attention network, specifically comprising: The node attention coefficient of a neighbor node is obtained through a self-attention mechanism, and a node representation is updated according to the node attention coefficient; wherein the update model of the node representation is: , wherein, is the feature representation of the updated node , is an activation function, represents a neighbor set of the node , is a parameter matrix, is the feature representation of a neighbor node .

3. The multi-view drug pair reaction prediction method of claim 1, wherein, obtaining feature information of a drug to be predicted and known reaction information between drugs, specifically comprising: Obtaining characteristic information of a drug to be predicted and known reaction information between drugs; and constructing a multi-relation network graph according to the characteristic information and the known reaction information ; wherein, , in the formula, a drug node, a reaction between drugs, is a characteristic of a drug; constructing an adjacency view according to the known reaction information, specifically comprising: Based on the multi-relationship network diagram Build adjacency view Wherein, in the adjacent view, if two drugs are known to interact, then in Set the corresponding position in the matrix to 1, otherwise set it to 0; constructing a network according to the known reaction information, and constructing a diffusion view by using personalized random walk PPR, specifically comprising: According to the multi-relation network graph , a diffusion view is constructed ; wherein the diffusion view is constructed according to a model: Wherein, (0,1] represents the instantaneous transition probability in random walk, is an identity matrix, is an adjacency matrix, is the degree matrix of the adjacency matrix .

4. The multi-view drug pair reaction prediction method of claim 3, wherein, constructing a drug feature matrix according to the feature information; and constructing a KNN view according to the drug feature matrix, specifically comprising: constructing an initial feature matrix of each feature type according to the feature information; performing Jaccard similarity calculation on each initial feature matrix to obtain a similarity matrix of each feature type; splicing the similarity matrices of each feature type to obtain a drug feature matrix; wherein the drug feature matrix contains feature vectors of each drug to be predicted; According to the drug feature matrix, cosine similarity between nodes is calculated, and K with the highest similarity is selected as neighbors of the node; wherein, the calculation model of the cosine similarity is: , wherein, represents the similarity between vectors , represents the feature vector of the drug , represents the feature vector of the drug , and the symbol represents the Euclidean norm operation;​ connecting an edge between a node and a neighbor, obtaining the KNN view .

5. The multi-view drug pair reaction prediction method of claim 1, wherein, fusing the three view-level embedding vectors to obtain an embedding representation of each drug, specifically comprising: The view attention coefficients of the updated view-level embedding vectors are obtained, and embedding representations of each drug are obtained by fusing the three updated view-level embedding vectors according to the view attention coefficients.

6. A multi-view drug pair reaction prediction device, characterized by, The multi-view drug pair reaction prediction method is suitable for executing the multi-view drug pair reaction prediction method. The multi-view drug pair reaction prediction device comprises: An initial information acquisition module is configured to acquire feature information of drugs to be predicted and known reaction information between drugs. A first view acquisition module is configured to construct an adjacency view according to the known reaction information. A second view acquisition module is configured to construct a network according to the known reaction information and construct a diffusion view by using personalized random walk PPR. A third view acquisition module is configured to construct a drug feature matrix according to the feature information. A view-level vector acquisition module is configured to acquire three view-level embedding vectors by deep fusion of the drug feature matrix and the adjacency view, the diffusion view, and the KNN view respectively through a node-level attention network. A fusion module is configured to fuse the three view-level embedding vectors by a view-level attention network to obtain embedding representations of each drug. A combination module is configured to combine the embedding representations of each drug to obtain a plurality of drug pair vectors, wherein the reaction type of part of the drug pair vectors is known. A training module is configured to train a classification model by taking drug pair vectors with known reaction types as a training set. A prediction module is configured to input drug pair vectors with unknown reaction types into the classification model to obtain a predicted reaction type. The computer program can be executed by the processor to implement the multi-view drug pair reaction prediction method.

7. A multi-view drug pair reaction prediction device characterized by comprising: The computer readable storage medium comprises a stored computer program, wherein when the computer program runs, the device where the computer readable storage medium is located executes the multi-view drug pair reaction prediction method.

8. A computer-readable storage medium, characterized in that, ​

Images