Multi-relation contrastive learning method for drug interaction prediction, system, medium and device

Abstract

The application discloses a multi-relation contrast learning drug interaction prediction method, system, medium and equipment, belongs to the technical field of graph neural networks and drug interaction prediction, and aims to solve the technical problem of how to capture implicit information about drug interaction and improve the accuracy of drug interaction prediction.The technical solution is as follows: obtaining SMILES sequences and molecular graph features of two drugs to be predicted, integrating multi-dimensional feature information of the two drugs to be predicted, and capturing comprehensive relationship features inside and between the two drugs to be predicted; converting the SMILES sequences of the two drugs to be predicted into drug molecular graphs by using an RDKit tool; extracting drug features from the drug molecular graphs by using a TrimNet method; obtaining a DDI network graph of the two drugs to be predicted and drugs with known interaction relationships; taking the feature vectors of the drug molecules as the feature vectors of the DDI network graph, fusing the drug molecular feature vectors and external DDI features by using a multi-layer GAT network layer, and obtaining updated drug representation vectors.

Description

Technical Field

[0001] This invention relates to the field of graph neural networks and drug interaction prediction technology, specifically to a method, system, medium, and device for predicting drug interactions using multi-relationship comparative learning. Background Technology

[0002] Drug-drug interactions (DDIs) refer to the combined effects of two or more drugs taken simultaneously or within a certain timeframe, meaning one drug may alter the effects of other drugs. With the increasing number of approved drugs on the market, combination therapy for complex diseases has become a common medical strategy. However, the concurrent use of multiple drugs can lead to adverse drug events (ADRs), meaning the efficacy of one drug may be altered by the presence of another. Effectively identifying potential DDIs can minimize unintended ADRs and maximize synergistic effects. Therefore, predicting potential drug interactions is crucial for physicians and patients to reduce the risk and cost of adverse drug events.

[0003] With the continuous enrichment and improvement of drug-related databases, deep learning methods have been widely applied in drug discovery index (DDI) prediction, greatly improving its performance. In early research, most DDI prediction models used chemical fingerprints as input. Chemical fingerprints are sequence data that can describe specific properties of drugs, such as chemical substructures, related targets, and side effects. In recent years, numerous researchers in the field of graph data structures have promoted the widespread adoption of graph data structure analysis techniques, and some widely used drug feature encoding methods rely on molecular graph structures. Drugs are referred to as chemical molecules with spatial structures. In drug molecular structure-based methods, the model input is a graph or data that can be converted into a graph, thus using adjacency matrices and feature matrices to represent the drug graph. While research on DDI prediction has achieved superior performance, most DDI prediction algorithms still have certain limitations. Due to the complexity of drug physicochemical properties, most models model information on a single view, considering either structural or interaction information. Furthermore, previous graph-based research lacked consideration of the importance of different neighbors, resulting in the correlation between node features not being well integrated into the model. In addition, there are some extremely rare DDI events, known as rare events. Because rare events involve extremely few labeled instances, they cannot provide sufficient supervision signals, making it challenging for most models to predict rare DDI events.

[0004] Therefore, how to capture implicit information about drug interactions and improve the accuracy of drug interaction prediction is a technical problem that urgently needs to be solved. Summary of the Invention

[0005] The technical objective of this invention is to provide a method, system, medium, and device for predicting drug interactions using multi-relational contrastive learning, in order to address the problem of how to capture implicit information about drug interactions and improve the accuracy of drug interaction prediction.

[0006] The technical objective of this invention is achieved as follows: a multi-relationship contrastive learning method for predicting drug interactions, the specific method of which is as follows:

[0007] The SMILES (Simplified Molecular-Input Line-Entry System) sequence and molecular graph features of the two drugs to be predicted are obtained, and the multidimensional feature information of the two drugs to be predicted is integrated to capture the comprehensive relationship features within and between the two drugs.

[0008] The RDKit tool was used to convert the SMILES sequences of the two drugs to be predicted into drug molecular graphs. Drug features were then extracted from the drug molecular graphs using the TrimNet method (an advanced variant of message passing neural networks). The drug molecular graphs are composed of atoms and chemical bonds.

[0009] Obtain the DDI network diagram (drug-drug interaction network diagram) of the two drugs to be predicted and the drugs with known interactions;

[0010] The feature vector of the drug molecule is used as the feature vector of the DDI network graph. The drug molecule feature vector and the external DDI features are fused using a multi-layer GAT network to obtain the updated drug representation vector.

[0011] Using a dual-view negative correspondence enhancement strategy, two views of the damaged graph are created by shuffling node features and edge features respectively, and then multi-relation comparison learning is achieved on the DDI network graph under the local-total information maximization strategy.

[0012] The drug molecule features and the updated drug molecule representation vector are concatenated and input into an interaction predictor (an interaction predictor is a model or tool used to predict the likelihood of interactions between different entities or objects. Such predictors are typically based on machine learning and data analysis techniques, building models by learning and analyzing interaction patterns in historical data, and used to predict future or new interactions) to obtain the interaction relationship between the two drugs to be predicted.

[0013] Preferably, the RDKit tool is used to convert the SMILES sequences of the two drugs to be predicted into drug molecular maps. The drug features are then extracted from the drug molecular maps using the TrimNet method (an advanced variant of a message-passing neural network), as follows:

[0014] Message Passing Phase: Define a message function, treating each drug molecule as a graph, each atom in the drug molecule as a node, and the chemical bonds connecting the atoms as edges; assign each atom in the molecule a randomly initialized embedding vector (considering its nuclear charge number) from a standard normal distribution. Specifically, the message passing phase includes T iterations, in each iteration, the hidden state associated with each atom i in each drug molecule graph is... The updated formula is as follows:

[0015]

[0016] in, Indicates the neighbors of atom i; This represents the message from atom j to atom i; U() represents the update function; then, neighborhood messages are aggregated through a layer-normalized gated recurrent unit (LN-GRU is a special type of recurrent neural network (RNN) unit that combines the advantages of layer normalization and gated recurrent units (GRU). This structure exhibits powerful performance when processing sequential data, especially time-dependent data such as audio and text). The message function is defined by a multi-head attention mechanism, as shown in the following formula:

[0017]

[0018] α ij =Softmax(σ(u) T [W s s i ∥∥W e e ij ∥∥W s s j ]));

[0019] Where ∥ represents vector concatenation; ⊙ represents element-wise multiplication; The hidden state of the edge between i and j is represented by σ; the LeakyReLU nonlinear function is represented by α. ij The scalar representing the measurement of attention; the vector u and the transformation matrix W of all features. s and W e These are learnable parameters;

[0020] Readout Phase: The TrimNet method uses a Set2Set architecture (Set2Set is a neural network architecture for processing set data, which can transform unordered set inputs into fixed-size vector representations, thereby encoding the set; the Set2Set architecture consists of two key parts: an encoding module and a decoding module. The encoding module is responsible for transforming each element in the set into a vector representation and aggregating these vectors into a fixed-size representation through a pooling mechanism (such as an attention mechanism). The decoding module is responsible for recovering the information of the original set from this representation or using this representation for further processing in downstream tasks) to read out the message function, summarizing all atomic embeddings after the message passing phase into drug features, represented as g. I .

[0021] As a preferred embodiment, the DDI network diagram (drug-drug interaction network diagram) is represented as follows:

[0022] Drug molecular map in, ε is the set of nodes, and ε is the set of edges; specifically, the i-th atom in the node set of drug I is represented as a node. For drug d I Represent the i-th atom in the node set as a node. The undirected chemical bond connecting the i-th and j-th atoms is denoted as e. ij ∈ε;

[0023] DDI Network Diagram in, This indicates the interaction pathways between drug molecules.

[0024] As a preferred approach, the feature vector of the drug molecule is used as the feature vector of the DDI network graph. A multi-layer GAT network is then used to fuse the drug molecule feature vector and external DDI features to obtain the updated drug representation vector, as detailed below:

[0025] In order to learn the drug representation vector in the DDI network, considering that the importance of different neighboring molecular nodes of drug molecules may be different, a self-attention mechanism is adopted to adaptively learn the weights, that is, a self-attention mechanism is implemented for each molecular node in the DDI network graph.

[0026] The softmax function is introduced to regularize all neighboring molecular nodes of a molecular node to obtain the updated attention coefficients. Specifically, for drug d... I Using the learned embedding vector g I As the feature vector of the GAT network; where g = {g1, g2, ..., g N}, Furthermore, by learning the interaction information between drug molecules in the DDI network diagram, the influence of neighbor J on drug I is calculated using the following formula:

[0027]

[0028] Where N represents all neighbors of node I; attention mechanism α IJ e is obtained by regularizing the neighboring node J of node I using the softmax function. IJ This makes the attention coefficient easier to calculate and compare; IJ The calculation formula is as follows:

[0029]

[0030] Where ∥ denotes the concatenation operation; T denotes the transpose; w denotes the parameter matrix; g I and g J Let represent the feature vectors of drug I and drug J, respectively; the attention mechanism is a single-layer feedforward neural network, with a LeakyReLu function added to the output layer of the feedforward neural network. LeakyReLu represents a function that assigns a non-zero slope to all negative values; similarly, through Obtain the attention scores of other adjacent molecular nodes;

[0031] The updated molecular I representation vector z is obtained by weighting the attention coefficients among the different molecular nodes after the update. I and the representation vector z of the molecular node J J The formula is as follows:

[0032]

[0033]

[0034] As a preferred approach, multi-relation contrastive learning refers to implementing multi-relation contrastive learning using a dual-view negative correspondence enhancement strategy under the mutual information maximization strategy. This involves creating two views of the damaged graph by shuffling node features and edge features. Then, multi-relation contrastive learning is implemented on the DDI network graph D under the local-global mutual information maximization strategy. Here, mutual information (MI) is a method used to measure the correlation between two random variables, that is, maximizing the mutual information between their inputs and outputs without focusing on their exact values.

[0035] More preferably, a dual-view negative correspondence enhancement strategy is used. By shuffling node features and edge features, two views of the damaged graph are created respectively. Then, multi-relation comparison learning is performed on the DDI network graph under the local-total information maximization strategy, as follows:

[0036] The dual-view negative correspondence enhancement strategy is used to create two views of the damaged graph by shuffling node features and edge features, and then the DDI network graph is implemented under the local-total information maximization strategy.

[0037] Destroy DDI network diagram D n and DDI network diagram D r To obtain a damage map and Two views, i.e., through Disrupting drug characteristics g and Disrupt edge features;

[0038] In the DDI network diagram D n and DDI network diagram D r Use the shared GAT encoder to obtain the corresponding "fake" drug representation. and

[0039] Given the original drug representation Z, use the readout function Γ to obtain the global representation represented by x = Γ(Z).

[0040] The training objective of contrastive learning is to maximize the consistency between Z and x. The difference between x and y is represented by the following loss function:

[0041]

[0042] Preferably, the drug molecule features and the updated drug molecule representation vector are concatenated and input into the interaction predictor to obtain the interaction relationship between the two drugs to be predicted, as follows:

[0043] Drug interaction prediction: For each drug pair (I,J), the characteristics g of each drug pair (I,J) have been obtained. I and g J And ultimately, z I and z J ;Characteristics g of each drug pair (I,J) I and g J And ultimately, z I and z J Connecting them to form drug pairs represents z (I,J) =z I ∥g I ∥z J ∥g J ;z (I,J) The data is input into a multilayer perceptron (MLP) and combined with a softmax function to obtain multi-class prediction results for drug pairs, as follows:

[0044]

[0045] in,

[0046] Loss function: The training objective for DDI event prediction is to minimize the loss function, as shown in the following formula:

[0047]

[0048] Where Ω represents the training set; This represents the predicted probability that the drug pair (I,J) belongs to relation type r; This represents the corresponding real label;

[0049] Optimize the loss function: Obtain the total loss by combining the loss function, as shown in the formula below:

[0050] l = l c +αl r +βl n ;

[0051] Here, α and β represent hyperparameters that balance the contributions of different tasks.

[0052] A multi-relationship contrastive learning drug interaction prediction system is provided to implement the multi-relationship contrastive learning drug interaction prediction method described above. The system includes:

[0053] The drug SMILES sequence acquisition module is used to acquire the SMILES (Simplified Molecular-Input Line-Entry System) sequence and molecular graph features of the two drugs to be predicted, integrate the multidimensional feature information of the two drugs to be predicted, and capture the comprehensive relationship features within and between the two drugs.

[0054] The drug molecule feature acquisition module is used to convert the SMILES sequences of two drugs to be predicted into drug molecule maps using the RDKit tool, and then extract drug features from the drug molecule maps using the TrimNet method (an advanced variant of message passing neural networks); wherein, the drug molecule map is composed of atoms and chemical bonds;

[0055] The DDI network diagram acquisition module is used to acquire the DDI network diagram (drug-drug interaction network diagram) of the two drugs to be predicted and the drugs with known interaction relationships.

[0056] The DDI network graph learning module is used to use the feature vector of the drug molecule as the feature vector of the DDI network graph. It uses a multi-layer GAT network layer to fuse the feature vector of the drug molecule and the external DDI features to obtain the updated drug representation vector.

[0057] The dual-view negative correspondence enhancement module is used to create two views of the damaged graph by shuffling node features and edge features, and then perform multi-relation comparison learning on the DDI network graph under the local-total information maximization strategy.

[0058] The interaction prediction module is used to connect drug molecule features and updated drug molecule representation vectors and input them into the interaction predictor to obtain the interaction relationship between the two drugs to be predicted.

[0059] An electronic device includes: a memory and at least one processor;

[0060] The memory contains computer programs;

[0061] The at least one processor executes the computer program stored in the memory, causing the at least one processor to perform the multi-relationship comparison learning drug interaction prediction method as described above.

[0062] A computer-readable storage medium storing a computer program that can be executed by a processor to implement the multi-relational contrastive learning drug interaction prediction method as described above.

[0063] The multi-relation comparison learning method, system, medium, and device of the present invention have the following advantages:

[0064] (I) This invention predicts drug interactions using MR-GAT, specifically by hierarchically integrating drug structure information from the drug molecule graph and drug interaction information from the DDI event graph. A multi-layer GAT network is used to process the fusion of drug features in each view and features from different neighbors in the external DDI network graph, thereby improving DDI event prediction. In MR-GAT, this invention uses the standard Simplified Molecular-Input Line-Entry for drug molecules. The system (SMILES) string and the DDI network connection matrix are used as inputs. Specifically, the first stage is the drug molecule feature extraction stage. The input drug SMILES sequence is used to extract drug molecule features g through the TrimNet method (an advanced variant of message passing neural networks). The feature vector g of the drug molecule is used as the feature vector of the DDI network graph. The drug molecule feature vector and the external DDI features are fused through the GAT encoder to obtain the updated drug representation vector z to update the drug embedding. Then, a dual-view negative correspondence enhancement strategy is used to create two views of the damaged graph by shuffling node features and edge features. Then, multi-relation contrastive learning is performed on the DDI network graph under the local-global mutual information maximization scheme to capture implicit information about rare events and achieve accurate prediction of DDI. Finally, there is an interaction predictor responsible for predicting missing interactions in the DDI network. The obtained drug molecule features are connected with the updated drug molecule representation vector to obtain the final interaction embedding and perform drug interaction prediction, which further improves the prediction accuracy.

[0065] (ii) In order to capture implicit information about rare events, this invention uses a dual-view negative correspondence enhancement strategy, which creates two views of the damaged graph by shuffling node features and edge features, and achieves multi-relation comparison learning of the DDI network graph under the local-global mutual information maximization scheme.

[0066] (III) Through extensive experiments on two datasets, this invention demonstrates that the MR-GAT provided by this invention has good generalization performance compared with various baseline methods. Attached Figure Description

[0067] The invention will be further described below with reference to the accompanying drawings.

[0068] Appendix Figure 1 This is a flowchart illustrating a multi-relational contrastive learning method for predicting drug interactions.

[0069] Appendix Figure 2 This is an example diagram of a multi-view structure. Detailed Implementation

[0070] The multi-relationship comparison learning drug interaction prediction method, system, medium, and device of the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0071] Example 1:

[0072] As attached Figure 1 As shown in the figure, this embodiment provides a multi-relationship contrastive learning method for predicting drug interactions, which is as follows:

[0073] S1. Obtain the SMILES (Simplified Molecular-Input Line-Entry System) sequence and molecular graph features of the two drugs to be predicted, integrate the multidimensional feature information of the two drugs to be predicted, and capture the comprehensive relationship features within and between the two drugs to be predicted.

[0074] S2. Use the RDKit tool to convert the SMILES sequences of the two drugs to be predicted into drug molecular maps, and extract drug features from the drug molecular maps using the TrimNet method (an advanced variant of message passing neural networks); the drug molecular maps are composed of atoms and chemical bonds.

[0075] S3. Obtain the DDI network diagram (drug-drug interaction network diagram) of the two drugs to be predicted and the drugs with known interaction relationships;

[0076] S4. Use the feature vector of the drug molecule as the feature vector of the DDI network graph, and use a multi-layer GAT network layer to fuse the feature vector of the drug molecule and the external DDI features to obtain the updated drug representation vector.

[0077] S5. Using a dual-view negative correspondence enhancement strategy, two views of the damaged graph are created by shuffling node features and edge features respectively, and then multi-relation comparison learning is achieved on the DDI network graph under the local-total information maximization strategy.

[0078] S6. Concatenate the drug molecule features and the updated drug molecule representation vector and input them into the interaction predictor (an interaction predictor is a model or tool used to predict the likelihood of interactions between different entities or objects. Such predictors are typically based on machine learning and data analysis techniques, building models by learning and analyzing interaction patterns in historical data, and using them to predict future or new interactions) to obtain the interaction relationship between the two drugs to be predicted.

[0079] In step S2 of this embodiment, the RDKit tool is used to convert the SMILES sequences of the two drugs to be predicted into drug molecule maps. The drug features are then extracted from the drug molecule maps using the TrimNet method (an advanced variant of a message-passing neural network) as follows:

[0080] S201, Message Passing Phase: Define a message function, treating each drug molecule as a graph, each atom in the drug molecule as a node, and the chemical bonds connecting the atoms as edges; assign each atom in the molecule a randomly initialized embedding vector (considering its nuclear charge number) from a standard normal distribution. Specifically, the message passing phase includes T iterations. In each iteration, the hidden state associated with each atom i in each drug molecule graph is... The updated formula is as follows:

[0081]

[0082] in, Indicates the neighbors of atom i; This represents the message from atom j to atom i; U() represents the update function; then, neighborhood messages are aggregated through a layer-normalized gated recurrent unit (LN-GRU is a special type of recurrent neural network (RNN) unit that combines the advantages of layer normalization and gated recurrent units (GRU). This structure exhibits powerful performance when processing sequential data, especially time-dependent data such as audio and text). The message function is defined by a multi-head attention mechanism, as shown in the following formula:

[0083]

[0084] α ij =Softmax(σ(u) T [W s s i ∥∥W e e ij ∥∥W s s j ]));

[0085] Where ∥ represents vector concatenation; ⊙ represents element-wise multiplication; The hidden state of the edge between i and j is represented by σ; the LeakyReLU nonlinear function is represented by α. ij The scalar representing the measurement of attention; the vector u and the transformation matrix W of all features. s and W eThese are learnable parameters;

[0086] S202, Readout Stage: The TrimNet method uses a Set2Set architecture (Set2Set is a neural network architecture for processing set data, which can transform unordered set inputs into fixed-size vector representations, thereby encoding the set; the Set2Set architecture consists of two key parts: an encoding module and a decoding module. The encoding module is responsible for transforming each element in the set into a vector representation and aggregating these vectors into a fixed-size representation through a pooling mechanism (such as an attention mechanism). The decoding module is responsible for recovering the information of the original set from this representation, or using this representation for further processing in downstream tasks) to read out the message function, summarizing all atomic embeddings after the message passing stage into drug features, represented as g. I .

[0087] As attached Figure 2 As shown, the DDI network diagram (drug-drug interaction network diagram) in this embodiment is represented as follows:

[0088] As attached Figure 2 (I) and Appendix Figure 2 (II) shows the drug molecule diagrams of drug A and drug B to be predicted, respectively. And DDI network diagram D. (Attached) Figure 2 In (I), the internal structures of drugs A and B are shown graphically, where a node (orange and blue nodes) represents an atom, and an edge (black straight line) represents a chemical bond; (See attached diagram). Figure 2 In (II), solid lines (orange and blue) and dashed lines (black) represent existing and potential interactions, with different colored solid lines representing different types of interactions.

[0089] Drug molecular map Depend on Indicates; among which, ε is the set of nodes, and ε is the set of edges; specifically, the i-th atom in the node set of drug I is represented as a node. For drug d I Represent the i-th atom in the node set as a node. The undirected chemical bond connecting the i-th and j-th atoms is denoted as e. ij ∈ε. DDI network graph D is composed of It means that among them This indicates the interaction pathways between drug molecules.

[0090] In step S4 of this embodiment, the feature vector of the drug molecule is used as the feature vector of the DDI network graph. A multi-layer GAT network is used to fuse the drug molecule feature vector and external DDI features to obtain the updated drug representation vector, as detailed below:

[0091] S401. In order to learn the representation vector of the drug in the DDI network, considering that the importance of different neighboring molecular nodes of the drug molecule may be different, a self-attention mechanism is adopted to adaptively learn the weights, that is, a self-attention mechanism is implemented for each molecular node in the DDI network graph.

[0092] S402. Introduce the softmax function to regularize all neighboring molecular nodes of a molecular node to obtain the updated attention coefficients. Specifically: for drug d... I Using the learned embedding vector g I As the feature vector of the GAT network; where g = {g1, g2, ..., g N}, Furthermore, by learning the interaction information between drug molecules in the DDI network diagram, the influence of neighbor J on drug I is calculated using the following formula:

[0093]

[0094] Where N represents all neighbors of node I; attention mechanism α IJ e is obtained by regularizing the neighboring node J of node I using the softmax function. IJ This makes the attention coefficient easier to calculate and compare; IJ The calculation formula is as follows:

[0095]

[0096] Where ∥ denotes the concatenation operation; T denotes the transpose; w denotes the parameter matrix; g I and g J Let represent the feature vectors of drug I and drug J, respectively; the attention mechanism is a single-layer feedforward neural network, with a LeakyReLu function added to the output layer of the feedforward neural network. LeakyReLu represents a function that assigns a non-zero slope to all negative values; similarly, through Obtain the attention scores of other adjacent molecular nodes;

[0097] S403. Weight the attention coefficients between the updated molecular nodes to obtain the updated molecular I representation vector z. I and the representation vector z of the molecular node J J The formula is as follows:

[0098]

[0099]

[0100] In this embodiment, multi-relation contrastive learning refers to implementing multi-relation contrastive learning using a dual-view negative correspondence enhancement strategy under the mutual information maximization strategy. Two views of the damaged graph are created by shuffling node features and edge features respectively. Then, multi-relation contrastive learning is implemented on the DDI network graph D under the local-global mutual information maximization strategy. Mutual information (MI) is a method used to measure the correlation between two random variables, that is, to maximize the mutual information between their inputs and outputs without focusing on their exact values.

[0101] In step S5 of this embodiment, the dual-view negative correspondence enhancement strategy is used to create two views of the damaged graph by shuffling node features and edge features. Then, multi-relation comparison learning is performed on the DDI network graph under the local-total information maximization strategy, as follows:

[0102] S501. Using a dual-view negative correspondence enhancement strategy, two views of the damaged graph are created by shuffling node features and edge features respectively, and then the DDI network graph is implemented under the local-total information maximization strategy.

[0103] S502, Destroy the DDI network diagram D n and DDI network diagram D r To obtain a damage map and Two views, i.e., through Disrupting drug characteristics g and Disrupt edge features;

[0104] S503, in the DDI network diagram D n and DDI network diagram D r Use the shared GAT encoder to obtain the corresponding "fake" drug representation. and

[0105] S504. Given the original drug representation Z, use the readout function Γ to obtain the global representation represented by x = Γ(Z).

[0106] S505, The training objective of contrastive learning is to maximize the consistency between Z and x, and The difference between x and y is represented by the following loss function:

[0107]

[0108] In this embodiment, step S6 involves concatenating the drug molecule features and the updated drug molecule representation vector and inputting them into the interaction predictor to obtain the interaction relationship between the two drugs to be predicted, as follows:

[0109] S601. Drug Interaction Prediction: For each drug pair (I,J), the characteristics g of each drug pair (I,J) have been obtained. I and g J And ultimately, z I and z J ;Characteristics g of each drug pair (I,J) I and g J And ultimately, z I and z J Connecting them to form drug pairs represents z (I,J) =z I ∥g I ∥z J ∥g J ;z (I,J) The data is input into a multilayer perceptron (MLP) and combined with a softmax function to obtain multi-class prediction results for drug pairs, as follows:

[0110]

[0111] in,

[0112] S602, Loss Function: The training objective for DDI event prediction is to minimize the loss function, as shown in the following formula:

[0113]

[0114] Where Ω represents the training set; This represents the predicted probability that the drug pair (I,J) belongs to relation type r; This represents the corresponding real label;

[0115] S603. Optimize the loss function: Obtain the total loss by combining the loss function, as shown in the following formula:

[0116] l = l c +αl r +βl n ;

[0117] Here, α and β represent hyperparameters that balance the contributions of different tasks.

[0118] To verify the effectiveness of this embodiment, the proposed model was evaluated using two datasets of different sizes: BinaryDDI and MultiDDI. The smaller dataset, BinaryDDI, contains a relatively small number of drugs, including 570 drugs, 65 DDI events, and 37,264 DDIs. The larger dataset, MultiDDI, contains 1,700 drugs and a total of 191,570 DDIs, involving 86 DDI events. Note that during data preprocessing, we removed SMILES sequences that could not be converted into molecular maps in the datasets. Detailed information about the datasets is shown in Table 1.

[0119] Table 1 Statistical information of the dataset

[0120] BinaryDDI 570 65 37,264 MultiDDI 1704 86 192,284

[0121] Specifically, the number of DDI instances involved in each DDI event is counted, called the event frequency, and then the DDI events are divided into five groups according to the frequency. Table 1 lists the proportion of events in each group to all events. It is worth mentioning that, as shown in group [1,10], many events have extremely limited labeled instances. We treat the events in group [1,10] as rare events to evaluate the performance of the prediction model on all events and rare events. Each drug is represented as a canonical SMILES string. Information on the event grouping of the dataset is shown in Table 2.

[0122] Table 2 shows the proportion of the five groups of events among all events.

[0123] BinaryDDI 20.0％ 21.5％ 24.6％ 15.4％ 18.5％ MultiDDI 5.8％ 21.0％ 11.6％ 14.0％ 47.6％

[0124] To illustrate the superiority of the model proposed in this embodiment, it is compared with several baseline methods:

[0125] DeepDDI is the first DDI event prediction method that uses drug structural similarity as input.

[0126] Construct deep neural networks to predict the types of drug-drug interactions.

[0127] SSI-DDI proposes a deep learning framework that considers substructure-substructure interactions, directly applies GATs to the drug molecule map, and combines the embedding of multiple GAT layers with a common concern mechanism to obtain drug pair predictions.

[0128] TrimNet learns drug features from drug molecule maps using TrimNet and inputs the tandem features of two drugs into an MLP to predict the drug pair type.

[0129] MUFFIN is a multi-scale feature fusion deep learning model for DDI event prediction that integrates drug features learned from drug molecule graphs with pre-trained knowledge graph embeddings.

[0130] GoGNN proposes a graph neural network within a graph that models molecular interaction networks as interaction graphs of molecular graphs. It uses a GNN to learn the molecular graph representation, which is then updated by another GNN deployed on the interaction graph for molecular interaction prediction.

[0131] • MRCGNN hierarchically integrates drug structure information from the drug molecule map and drug interaction information from the DDI event map to improve DDI event prediction.

[0132] The experimental results are shown in Table 3:

[0133] Table 3. Results of the drug interaction prediction experiment

[0134]

[0135] The results of the drug interaction prediction experiments are shown in Table 3. It can be seen that the MR-GAT method of this embodiment achieved excellent results on both datasets:

[0136] (1) Compared with SSI-DDI, TrimNet and MUFFIN, which only consider drug structure information, MR-GAT improves accuracy by 11.77%, 4.73% and 7.74% on small datasets, respectively, and by 5.93%, 2.48% and 0.91% on large datasets, respectively.

[0137] (2) Compared with DeepDDI, which only considers drug interaction information, MR-GAT outperforms by 12.36% on small datasets and by 2.78% on large datasets, which means that considering drug molecule graphs is also beneficial for DDI event prediction.

[0138] (3) Among all baselines, MRCGNN showed better performance than most other baselines, which means the advantage of integrating information from both drug molecule maps and DDI event maps.

[0139] On two datasets of different sizes, this embodiment still achieves an improvement in accuracy compared to MRCGNN. This may be because MR-GAT can not only effectively integrate drug structure information and drug interactions, but also capture the implicit information behind the multi-relation DDI event graph through multi-relation contrastive learning, and consider the importance of different neighbor nodes of drug molecules and the information balance between multiple views, thus learning a more comprehensive drug representation.

[0140] Example 2:

[0141] This embodiment provides a multi-relation comparison learning drug interaction prediction system, which is used to implement the multi-relation comparison learning drug interaction prediction method in Embodiment 1; the system includes:

[0142] The drug SMILES sequence acquisition module is used to acquire the SMILES (Simplified Molecular-Input Line-Entry System) sequence and molecular graph features of the two drugs to be predicted, integrate the multidimensional feature information of the two drugs to be predicted, and capture the comprehensive relationship features within and between the two drugs.

[0143] The drug molecule feature acquisition module is used to convert the SMILES sequences of two drugs to be predicted into drug molecule maps using the RDKit tool, and then extract drug features from the drug molecule maps using the TrimNet method (an advanced variant of message passing neural networks); wherein, the drug molecule map is composed of atoms and chemical bonds;

[0144] The DDI network diagram acquisition module is used to acquire the DDI network diagram (drug-drug interaction network diagram) of the two drugs to be predicted and the drugs with known interaction relationships.

[0145] The DDI network graph learning module is used to use the feature vector of the drug molecule as the feature vector of the DDI network graph. It uses a multi-layer GAT network layer to fuse the feature vector of the drug molecule and the external DDI features to obtain the updated drug representation vector.

[0146] The dual-view negative correspondence enhancement module is used to create two views of the damaged graph by shuffling node features and edge features, and then perform multi-relation comparison learning on the DDI network graph under the local-total information maximization strategy.

[0147] The interaction prediction module is used to connect drug molecule features and updated drug molecule representation vectors and input them into the interaction predictor to obtain the interaction relationship between the two drugs to be predicted.

[0148] Example 3:

[0149] This embodiment also provides an electronic device, including: a memory and a processor;

[0150] The memory stores the instructions executed by the computer.

[0151] The processor executes computer execution instructions stored in the memory, causing the processor to perform the multi-relationship contrastive learning drug interaction prediction method in any embodiment of the present invention.

[0152] The processor can be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), off-the-shelf programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The processor can be a microprocessor or any conventional processor.

[0153] Memory is used to store computer programs and / or modules. The processor implements various functions of the electronic device by running or executing the computer programs and / or modules stored in the memory, and by accessing data stored in the memory. Memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, at least one application program required for a function, etc.; the data storage area can store data created based on the use of the terminal, etc. In addition, memory can also include high-speed random access memory, and can also include non-volatile memory, such as hard disks, RAM, plug-in hard disks, smart memory cards (SMC), secure digital cards (SD cards), flash memory cards, at least one disk storage device, flash memory devices, or other volatile solid-state storage devices.

[0154] Example 4:

[0155] This embodiment also provides a computer-readable storage medium storing multiple instructions, which are loaded by a processor to cause the processor to execute the multi-relationship contrastive learning drug interaction prediction method of any embodiment of the present invention. Specifically, a system or apparatus equipped with a storage medium may be provided, on which software program code implementing the functions of any of the above embodiments is stored, and the computer (or CPU or MPU) of the system or apparatus may read and execute the program code stored in the storage medium.

[0156] In this case, the program code read from the storage medium can itself implement the function of any of the above embodiments, and therefore the program code and the storage medium storing the program code constitute part of the present invention.

[0157] Storage media embodiments for providing program code include floppy disks, hard disks, magneto-optical disks, optical disks (such as CD-ROM, CD-R, CD-RW, DVD-ROM, DVD-RYM, DVD-RW, DVD+RW), magnetic tapes, non-volatile memory cards, and ROMs. Alternatively, program code can be downloaded from a server computer via a communication network.

[0158] Furthermore, it should be clear that not only can the program code read by the computer be executed, but also the operating system or other components operating on the computer can be instructed based on the program code to perform some or all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0159] Furthermore, it is understood that the program code read from the storage medium is written to the memory set in the expansion board inserted into the computer or to the memory set in the expansion unit connected to the computer. Then, based on the instructions of the program code, the CPU or other components installed on the expansion board or expansion unit execute some and all of the actual operations, thereby realizing the function of any of the embodiments described above.

[0160] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A multi-relational contrastive learning method for predicting drug interactions, characterized in that, The method is as follows: The study obtains the SMILES sequence and molecular graph features of the two drugs to be predicted, integrates the multidimensional feature information of the two drugs to be predicted, and captures the comprehensive relationship features within and between the two drugs. The RDKit tool was used to convert the SMILES sequences of the two drugs to be predicted into drug molecular maps, and the drug features were extracted from the drug molecular maps using the TrimNet method; the drug molecular maps are composed of atoms and chemical bonds. Obtain the DDI network diagram of the two drugs to be predicted and the drugs with known interactions; The feature vector of the drug molecule is used as the feature vector of the DDI network graph. The drug molecule feature vector and the external DDI features are fused using a multi-layer GAT network to obtain the updated drug representation vector. Using a dual-view negative correspondence enhancement strategy, two views of the damaged graph are created by shuffling node features and edge features respectively, and then multi-relation comparison learning is achieved on the DDI network graph under the local-total information maximization strategy. The drug molecule features and the updated drug molecule representation vector are concatenated and input into the interaction predictor to obtain the interaction relationship between the two drugs to be predicted. Specifically, the feature vector of the drug molecule is used as the feature vector of the DDI network graph. A multi-layer GAT network is used to fuse the drug molecule feature vector and the external DDI features to obtain the updated drug representation vector, as follows: The weights are adaptively learned using a self-attention mechanism, that is, a self-attention mechanism is implemented for each molecular node in the DDI network graph; The softmax function is introduced to regularize all neighboring molecular nodes of a molecular node to obtain the updated attention coefficients. Specifically, for drugs... Using the learned embedding vectors As the feature vector of the GAT network; where... This allows for the learning of interactions between drug molecules in the DDI network diagram, including neighboring molecules. For drugs The formula for calculating the impact is as follows: ； in, Represents a node All neighbors; attention mechanism The softmax function is used to evaluate the nodes. neighboring nodes After regularization get; The calculation formula is as follows: ； in, Indicates a serial operation; Indicates transpose; Represents the parameter matrix; and They represent drugs and medicine The feature vector; the attention mechanism is a single-layer feedforward neural network, with an addition added to the output layer of the feedforward neural network. LeakyReLu function, LeakyReLu To represent a function, we assign a non-zero slope to all negative values; similarly, by... Obtain the attention scores of other adjacent molecular nodes; The updated molecular nodes are obtained by weighting the attention coefficients between them. Representing vectors and molecular nodes J The representation vector The formula is as follows: ； ； Multi-relation contrastive learning refers to the implementation of multi-relation contrastive learning using a dual-view negative correspondence enhancement strategy under the mutual information maximization strategy. This involves creating two views of a corrupted graph by shuffling node and edge features. Then, multi-relation contrastive learning is performed on the DDI network graph D under the local-global mutual information maximization strategy. Mutual information is a method used to measure the correlation between two random variables, i.e., maximizing the mutual information between their inputs and outputs.

2. The multi-relationship contrastive learning drug interaction prediction method according to claim 1, characterized in that, The RDKit tool was used to convert the SMILES sequences of the two drugs to be predicted into drug molecular maps. The drug features were then extracted from the drug molecular maps using the TrimNet method, as follows: Message Passing Phase: Define a message function that treats each drug molecule as a graph, each atom in the drug molecule as a node, and the chemical bonds connecting the atoms as edges. Assign each atom in the molecule a randomly initialized embedding vector from a standard normal distribution. Specifically, the message passing phase consists of T iterations. In each iteration, the embedding vector is used to embed each atom in the graph of each drug molecule. Related hidden state The updated formula is as follows: ； in, Represents atoms Neighbors; Represents atoms To the atom The message; The update function is represented; then, neighborhood messages are aggregated through layer-normalized gated recurrent units; the message function is defined by a multi-head attention mechanism, as shown in the following formula: ； ； in, This represents the vector concatenation operation; Represents element-wise product; express and The hidden state of the edges between them; Represents the LeakyReLU nonlinear function; A scalar representing the measurement of attention; a vector and all characteristic transformation matrices and These are learnable parameters; Readout Phase: The TrimNet method uses a Set2Set architecture to read out message functions, summarizing all atomic embeddings after the message passing phase into drug features, represented as follows: .

3. The multi-relationship contrastive learning method for predicting drug interactions according to claim 1, characterized in that, The specific representation of a DDI network diagram is as follows: Drug molecular map ;in, It is a set of nodes. It is an edge set; specifically: the drug I The first node in the node set Each atom is represented as a node. For drugs Set the nodes in the first place Each atom is represented as a node. , No. and the The undirected chemical bonds connecting atoms are represented as follows: ; DDI Network Diagram ;in, This indicates the interaction pathways between drug molecules.

4. The multi-relationship contrastive learning drug interaction prediction method according to claim 1, characterized in that, A dual-view negative correspondence enhancement strategy is used to create two views of the damaged graph by shuffling node and edge features. Then, multi-relation comparison learning is performed on the DDI network graph under the local-total information maximization strategy, as detailed below: The dual-view negative correspondence enhancement strategy is used to create two views of the damaged graph by shuffling node features and edge features, and then the DDI network graph is implemented under the local-total information maximization strategy. Destroy DDI network diagram and DDI network diagram To obtain a damage map and Two views, i.e., through Disrupt drug characteristics as well as Disrupt edge features; In the DDI network diagram and DDI network diagram Use the shared GAT encoder to obtain the corresponding "fake" drug representation. and ; Given the original drug representation Use the read function Γ to obtain the data from the read function Γ. Global representation ; The training objective of contrastive learning is to maximize and Consistency between them and and The difference between them is represented by the following loss function: ； 。 5. The multi-relationship contrastive learning method for predicting drug interactions according to claim 1, characterized in that, The drug molecule features and the updated drug molecule representation vector are concatenated and input into the interaction predictor to obtain the interaction relationship between the two drugs to be predicted, as follows: Drug interaction prediction: For each drug pair Each drug pair has been obtained Features and And the final expression and ; each drug pair Features and And the final expression and Connect them to form drug pairs ; The data is input into a multilayer perceptron and combined with the Softmax function to obtain multi-class prediction results for drug pairs, as follows: ； in, ; Loss function: The training objective for DDI event prediction is to minimize the loss function, as shown in the following formula: ； Where Ω represents the training set; Indicates the drug's effect Belongs to relation type The predicted probability; This represents the corresponding real label; Optimize the loss function: Obtain the total loss by combining the loss function, as shown in the formula below: ； in, and This represents the hyperparameters that balance the contributions of different tasks.

6. A multi-relational contrastive learning drug interaction prediction system, characterized in that, This system is used to implement the multi-relationship contrastive learning drug interaction prediction method as described in any one of claims 1 to 5; the system comprises: The drug SMILES sequence acquisition module is used to acquire the SMILES sequences and molecular graph features of the two drugs to be predicted, integrate the multidimensional feature information of the two drugs to be predicted, and capture the comprehensive relationship features within and between the two drugs. The drug molecule feature acquisition module is used to convert the SMILES sequences of two drugs to be predicted into drug molecule maps using the RDKit tool, and then extract drug features from the drug molecule maps using the TrimNet method; the drug molecule map is composed of atoms and chemical bonds; The DDI network diagram acquisition module is used to acquire the DDI network diagrams of the two drugs to be predicted and the drugs with known interaction relationships. The DDI network graph learning module is used to use the feature vector of the drug molecule as the feature vector of the DDI network graph. It uses a multi-layer GAT network layer to fuse the feature vector of the drug molecule and the external DDI features to obtain the updated drug representation vector. The dual-view negative correspondence enhancement module is used to create two views of the damaged graph by shuffling node features and edge features, and then perform multi-relation comparison learning on the DDI network graph under the local-total information maximization strategy. The interaction prediction module is used to connect drug molecule features and updated drug molecule representation vectors and input them into the interaction predictor to obtain the interaction relationship between the two drugs to be predicted.

7. An electronic device, characterized in that, include: Memory and at least one processor; The memory contains computer programs; The at least one processor executes the computer program stored in the memory, causing the at least one processor to perform the multi-relationship contrastive learning drug interaction prediction method as described in any one of claims 1 to 5.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that can be executed by a processor to implement the multi-relationship contrastive learning drug interaction prediction method as described in any one of claims 1 to 5.

Images