A deep learning-based drug-drug adverse reaction prediction method

Abstract

The application discloses a drug-drug adverse reaction prediction method based on deep learning, which comprises four steps: step one, using a message passing network with an attention mechanism to encode drug structure information; step two, using a similarity function and an automatic encoder to encode drug binding proteins; step three, splicing the different characteristics of the above-mentioned drugs into an adaptive learning module inspired by a meta-path to adaptively learn a meta-path subgraph and perform graph convolution operation to obtain drug characteristics; step four, sending the drug characteristics into a multilayer perceptron to predict drug-drug adverse reactions. The application proposes a method that can effectively capture drug-drug adverse reaction characteristics, solves the problem that common models directly use graph neural networks on interaction networks, and easily mixes noise, and through adaptive learning of effective feature information, the drug-drug adverse reaction prediction effect is improved.

Description

Technical Field

[0001] This invention relates to the fields of drug design and unstructured data feature extraction, and in particular to a method for extracting effective correlation information by fusing drug features at different levels and combining meta-paths for adaptive learning. Background Technology

[0002] Drug discovery is a field of bioinformatics. In the process of new drug development, effectively identifying adverse drug reactions is crucial not only for the pharmaceutical industry but also for patient health. Severe adverse reactions can cause irreversible damage to patients, leading to significant drug safety issues. Effectively identifying adverse drug reactions is essential to further protect patient health.

[0003] In recent years, against the backdrop of computer-aided drug research, an increasing number of machine learning methods have been successfully applied to drug-drug adverse reaction prediction. With the accumulation of biomedical data, machine learning methods can automatically learn features from rich datasets, transforming simple features into complex and effective ones. These effective features can then be used to more accurately predict adverse drug reactions, avoiding the tedious, time-consuming, and costly drawbacks of clinical trials, which is crucial for patients and society. Summary of the Invention

[0004] The purpose of this invention is to propose a method for predicting drug-drug adverse reactions by integrating different levels of drug characteristics with meta-pathway-inspired adaptive learning, based on the molecular structure information of drugs and information on binding proteins (transporters, carrier proteins, enzyme proteins, and target proteins). This method can significantly improve the accuracy of predicting adverse drug reactions.

[0005] The specific technical solution for achieving the objective of this invention is as follows:

[0006] A deep learning-based drug-adverse drug reaction prediction method is characterized by adaptively learning meta-path subgraphs of similar drug pairs and the interactions between adverse drug reactions to capture effective features for prediction. The method includes: Step 1: Encoding drug molecular structure information.

[0007] 1.1) The SMILES sequence of the drug is converted into a molecular graph using RDkit software, i.e., an undirected graph is constructed with atoms as nodes and the bonds between atoms as edges; for drug i, an undirected graph G is constructed. i (V,E) represents the molecular diagram of drug i, where V is the set of all atoms in the molecular diagram and E is the set of all bonds.

[0008] 1.2) Using message-passing networks with attention mechanisms to process undirected graphs Gi Encoding involves two phases: the message passing phase and the graph reading phase.

[0009] 1.3) The message passing phase consists of message functions and update functions. This phase runs for T time steps, first initializing the characteristics of each node. At this point, the time step t = 0. Then, the feature vectors of node v's neighboring nodes are aggregated using the message function, with the following formula:

[0010]

[0011] Where N(v) is the set of neighboring nodes of node v. It is a learnable parameter matrix. It is the feature vector of node v's neighbor node w;

[0012] Then, by combining two nonlinear transformation functions F and C as the update function, the neighbor aggregation features of node v are obtained. Features of node v The fusion update yields the feature vector of node v at time step (t+1). The formula is:

[0013]

[0014]

[0015] in The symbol represents a concatenation operation, ⊙ represents an inner product operation, and W represents a concatenation operation. c and W f These are two learnable parameters;

[0016] 1.4) After running for T time steps during the message passing phase, the final feature representation of each node in the drug molecule graph is obtained. Then, during the graph readout stage, the entire drug molecule graph is calculated as an eigenvector E. i ′ As structural information features of drug molecules; calculations are performed using a graph readout function with an effective attention mechanism, the formula is as follows:

[0017]

[0018] Where σ1 and π2 represent the sigmoid and tanh activation functions, respectively, W i1 and W i2 These are two learnable weight parameters, b i1 and b i2 These are two learnable bias vectors. It will be used as an attention mechanism to represent the importance score of each node;

[0019] Step 2: Encoding drug-binding protein information

[0020] 2.1) Drug-binding proteins are proteins that bind to drugs and influence their function to exert their effects. These include transport proteins, carrier proteins, enzyme proteins, and target proteins. Each binding protein has its own characteristics. First, a similarity function is used to calculate the characteristics of each binding protein. The formula is as follows:

[0021]

[0022] Where x i With x j Let i and j represent the feature vectors of drugs i and j on a certain binding protein, respectively; in this way, we can obtain the similarity matrix of transport protein, carrier protein, enzyme protein, and target protein.

[0023] 2.2) Then, the feature matrices of the four binding proteins are concatenated and fed into an autoencoder with an attention mechanism for encoding. The autoencoder is an unsupervised neural network model, which consists of two parts: an encoder and a decoder, and can learn the features of the input data; ultimately, it can obtain the binding protein features of each drug.

[0024] Step 3: Adaptive Learning

[0025] 3.1) The drug molecule structure information features and drug-bound protein features are concatenated and fed into the adaptive learning module. In the adaptive learning module, N+2 adjacency matrices are first formed, where N represents the number of adjacency matrices corresponding to each adverse reaction type. Each adjacency matrix is ​​formed by connecting drug pairs with interactions in an adverse reaction event. The two adjacency matrices are represented by the drug molecule structure information similarity feature map and the drug-bound protein similarity feature map. The similarity feature map is obtained through metric learning, with the following formula:

[0026]

[0027]

[0028] Where S FS [i,j] represents an adjacency matrix, where i and j correspond to drug i and drug j in the adjacency matrix, respectively. Φ FS It is a cosine similarity function with K weight parameters, E′ i and E′ j Features corresponding to drug i and drug j respectively, ∈ is the threshold, W kThese are learnable weight parameters; thus, drug molecule structure information similarity feature maps and drug-bound protein similarity feature maps are constructed by using the similarity of drug molecule structure information features and the similarity of drug-bound protein features.

[0029] 3.2) The obtained N+2 adjacency matrices are subjected to adaptive transformation. The adaptive transformation is achieved by adaptively learning the most effective meta-paths associated with drug-adverse drug reactions: firstly, a learnable weight matrix W is used. ψ Multiplying and summing the results with the N+2 adjacency matrices yields a subgraph, where the weight parameters α are learned from each of the N+2 adjacency matrices. t1 ∈W ψ , where α t1 For 1x1 weight parameters, α t1 A larger value indicates that the corresponding adjacency matrix is ​​more important to the prediction result, thus integrating all important adjacency matrix information into a single subgraph. The formula is:

[0030] A1=∑ t1∈T α t1 A t1 ;

[0031] 3.3) Step 3.2) is a 1-layer adaptive transformation, which is the subgraph formed by connecting neighboring nodes with a distance of 1 hop. The 1-layer adaptive transformation is then multiplied by matrix to obtain the meta-path of length l, which is the subgraph formed by connecting neighboring nodes with a distance of l hops. The formula is:

[0032] A l =(∑ t1∈T α t1 A t1 )(∑ t2∈T α t2 A t2 )…(∑ tl∈T α tl A tl ),

[0033] This adaptive learning produces a subgraph with an l-hop distance, which can fuse effective drug-adverse drug reaction type associations, where l is a constant.

[0034] 3.4) Multi-channel learning is adopted, that is, num-channel channels are set for learning, and finally num_channel subgraphs with l-hop distances are learned;

[0035] 3.5) The graph convolutional neural network is used to train and learn on the subgraphs with l-hop distances in num_channel. The nodes in each channel subgraph will fuse their neighbor information to form the final node feature matrix. The node feature matrices of multiple channels will be stacked to form the final feature matrix of the drug.

[0036] Step 4: Multilayer Perceptron Prediction

[0037] 4.1) Extract features of drug pairs that will cause adverse drug reactions from the final feature matrix, concatenate the features of the two drugs, and then feed them into a perceptron composed of two fully connected neural networks for learning. The formula is as follows:

[0038]

[0039]

[0040] Where W3 and W4 represent learnable weight parameters, b3 and b4 are bias vectors, and σ is the sigmoid activation function. This is the final predicted value. The model uses cross-entropy as the loss function for training and learning.

[0041] The technical concept of this invention is as follows: By encoding drug structure information and drug-binding protein information, features at different levels of the drug can be obtained. Fusing these two types of features provides a better representation of drug characteristics. This data is then fed into an adaptive learning module inspired by meta-paths. A similarity graph of the two features is constructed and added to the adverse reaction adjacency matrix, followed by adaptive learning of a new effective meta-path subgraph. The model utilizes the similarity principle in adverse drug reactions (ADRs), where similar drug pairs can lead to similar ADRs, and multiple ADRs can interact with each other. The model does not require pre-defined effective meta-paths; instead, it adaptively learns new meta-path subgraphs of similar drug pairs and the interactions between ADRs. Graph convolution operations are then used to fuse neighbor node information across multiple hops in the new subgraph, thereby more effectively predicting drug-adverse reaction types. This invention effectively captures drug-adverse reaction features by fusing feature information directly from nodes to neighbor nodes connected by a distance of l hops. This solves the problem of noise easily introduced into conventional models using graph neural networks directly on the interaction network. Furthermore, by adaptively learning effective associated feature information, it improves the prediction performance of drug-adverse reaction. Attached Figure Description

[0042] Figure 1 This is a diagram of a message passing network;

[0043] Figure 2Flowchart for encoding drug-binding protein information;

[0044] Figure 3 This is a flowchart of the present invention. Detailed Implementation

[0045] To facilitate understanding of the present invention, the invention will now be described in detail with reference to the accompanying drawings and embodiments. Figures 1-3 This embodiment divides the dataset into training, validation, and test sets in an 8:1:1 ratio. The method includes the following steps:

[0046] 1) See Figure 1 A schematic diagram of drug molecular structure information encoding, specifically including:

[0047] 1.1) The SMILES sequence of the drug is converted into a molecular graph using RDkit software. This involves constructing an undirected graph with atoms as nodes and the bonds between atoms as edges. For drug i, an undirected graph G is constructed. i (V,E) represents the molecular diagram of drug i, where V is the set of all atoms in the molecular diagram and E is the set of all bonds.

[0048] 1.2) Using message-passing networks with attention mechanisms to process undirected graphs G i Encoding involves two phases: the message passing phase and the graph reading phase.

[0049] 1.3) The message passing phase consists of message functions and update functions, and runs for three time steps. First, the characteristics of each node are initialized. At this point, the time step t = 0. Then, the feature vectors of node v's neighboring nodes are aggregated using the message function, with the following formula:

[0050]

[0051] Where N(v) is the set of neighboring nodes of node v. It is a learnable parameter matrix. It is the feature vector of node v's neighbor node w;

[0052] Then, by combining two nonlinear transformation functions F and C as the update function, the neighbor aggregation features of node v are obtained. Features of node v The fusion update yields the feature vector of node v at time step (t+1). The formula is:

[0053]

[0054]

[0055] in The symbol represents a concatenation operation, ⊙ represents an inner product operation, and W represents a concatenation operation. c and W f These are two learnable parameters;

[0056] 1.4) After running for 3 time steps in the message passing phase, the final feature representation of each node in the drug molecule graph is obtained. Then, during the graph readout stage, the entire drug molecule graph is calculated as an eigenvector E′. i As structural information features of drug molecules; calculations are performed using a graph readout function with an effective attention mechanism, the formula is as follows:

[0057]

[0058] Where σ1 and σ2 represent the sigmoid and tanh activation functions, respectively, W i1 and W i2 These are two learnable weight parameters, b i1 and b i2 These are two learnable bias vectors. It will be used as an attention mechanism to represent the importance score of each node;

[0059] 2) See Figure 2 Drug-binding protein information encoding, specifically including:

[0060] 2.1) Drug-binding proteins are proteins that bind to drugs and influence their function to exert their effects. These include transport proteins, carrier proteins, enzyme proteins, and target proteins. Each binding protein has its own characteristics. First, a similarity function is used to calculate the characteristics of each binding protein. The formula is as follows:

[0061]

[0062] Where x i With x j Let i and j represent the feature vectors of drugs i and j on a certain binding protein, respectively; in this way, we can obtain the similarity matrix of transport protein, carrier protein, enzyme protein, and target protein.

[0063] 2.2) Then, the feature matrices of the four binding proteins are concatenated and fed into an autoencoder with an attention mechanism for encoding. The autoencoder is an unsupervised neural network model, which consists of two parts: an encoder and a decoder, and can learn the features of the input data; ultimately, it can obtain the binding protein features of each drug.

[0064] 3) See Figure 3 Adaptive learning in [the context of learning] specifically includes:

[0065] 3.1) The drug molecule structure information features and drug-bound protein features are concatenated and fed into the adaptive learning module. In the adaptive learning module, 65+2 adjacency matrices are first formed, where 65 represents the number of adjacency matrices corresponding to each adverse reaction type. Each adjacency matrix is ​​formed by connecting drug pairs with interactions in an adverse reaction event. The two adjacency matrices are represented by the drug molecule structure information similarity feature map and the drug-bound protein similarity feature map. The similarity feature map is obtained through metric learning, with the following formula:

[0066]

[0067]

[0068] Where S FS [i,j] represents an adjacency matrix, where i and j correspond to drug i and drug j in the adjacency matrix, respectively. Φ FS It is a cosine similarity function with K weight parameters, E′ i and E′ j Features corresponding to drug i and drug j respectively, ∈ is the threshold, W k These are learnable weight parameters; thus, drug molecule structure information similarity feature maps and drug-bound protein similarity feature maps are constructed by using the similarity of drug molecule structure information features and the similarity of drug-bound protein features.

[0069] 3.2) The obtained 65+2 adjacency matrices are subjected to adaptive transformation. The adaptive transformation is achieved by adaptively learning the most effective meta-paths associated with drug-adverse drug reactions: firstly, a learnable weight matrix W is used. ψ Multiplying and summing the results with 65+2 adjacency matrices yields a subgraph, where the weight parameter α is learned from each of the 65+2 adjacency matrices. t1 ∈W ψ , where α t1 For 1x1 weight parameters, α t1 A larger value indicates that the corresponding adjacency matrix is ​​more important to the prediction result, thus integrating all important adjacency matrix information into a single subgraph. The formula is:

[0070] A1=∑ t1∈T α t1 A t1 ;

[0071] 3.3) Step 3.2) is a single-layer adaptive transformation, corresponding to a subgraph with a 1-hop distance. The model requires three layers of adaptive transformation and matrix multiplication to obtain a meta-path of length 3, which is the subgraph formed by connecting neighboring nodes with a 3-hop distance to a node, thus forming a multi-hop subgraph. The formula is:

[0072] A l =(∑ t1∈T α t1 A t1 )(∑ t2∈T α t2 A t2 )…(∑ tl∈T α tl A tl ),

[0073] In this way, the adaptive module can learn a subgraph with an l-hop distance, and can fuse effective drug-adverse drug reaction type association information;

[0074] 3.4) Multi-channel learning is adopted, that is, three channels are set for learning, and three subgraphs with a distance of three hops will be learned in the end;

[0075] 3.5) The graph convolutional neural network is used for training and learning on num_channel subgraphs with a 3-hop distance. The nodes on each channel subgraph will fuse their neighbor information to form the final node feature matrix. The node feature matrices of multiple channels will be stacked to form the final feature matrix of the drug.

[0076] 4) Multilayer perceptron prediction, specifically including:

[0077] 4.1) Extract features of drug pairs that will cause adverse drug reactions from the final feature matrix, concatenate the features of the two drugs, and then feed them into a perceptron composed of two fully connected neural networks for learning. The formula is as follows:

[0078]

[0079]

[0080] Where W3 and W4 represent learnable weight parameters, b3 and b4 are bias vectors, and σ is the sigmoid activation function. This is the final predicted value. The model uses cross-entropy as the loss function for training and learning.

Claims

1. A deep learning-based method for predicting adverse drug reactions, characterized by the following specific steps: Step 1: Encoding drug molecular structure information 1.1) The SMILES sequence of the drug was converted into a molecular graph using RDkit software, i.e., an undirected graph was constructed with atoms as nodes and the bonds between atoms as edges; for drugs Construct an undirected graph To indicate drugs Molecular diagram, in which It is the set of all atoms in the molecular diagram. It is the set of all keys; 1.2) Using message-passing networks with attention mechanisms to process undirected graphs Encoding involves two phases: the message passing phase and the graph reading phase. 1.3) The message passing phase consists of message functions and update functions, and execution takes place during this phase. At each time step, the features of each node are first initialized as follows: The time step at this time Then use message functions to interact with the nodes. The aggregated feature vector of neighboring nodes is calculated using the following formula: ， in It is a node The set of neighboring nodes, It is a learnable parameter matrix. It is a node neighboring nodes eigenvectors; Then use two nonlinear transformation functions and Combining the results as an update function will yield the nodes. Neighbor aggregation features With nodes Features Nodes are obtained through fusion and update. exist Feature vectors at time step Its formula is: , ， in This indicates a splicing operation. This represents the inner product operation. and These are two learnable parameters; 1.4) Run during the message passing phase After a certain time step, the final feature representation of each node in the drug molecule graph is obtained. Then, during the graph readout stage, the entire drug molecule graph is calculated as a feature vector. As structural information features of drug molecules; calculations are performed using a graph readout function with an effective attention mechanism, the formula is as follows: ， in and These represent the sigmoid and tanh activation functions, respectively. and These are two learnable weight parameters. and These are two learnable bias vectors. It will be used as an attention mechanism to represent the importance score of each node; Step 2: Encoding drug-binding protein information 2.1) Drug-binding proteins are proteins that bind to drugs and affect their function to exert their efficacy, including: Transport proteins, carrier proteins, enzyme proteins, and target proteins; binding proteins each have their own characteristics, firstly using similarity functions. The features of each binding protein are calculated using the following formula: ， in and They represent drugs and The feature vector on a certain binding protein; in this way, the similarity matrix of transport protein, carrier protein, enzyme protein, and target protein can be obtained. 2.2) Then, the feature matrices of the four binding proteins are concatenated and fed into an autoencoder with an attention mechanism for encoding. The autoencoder is an unsupervised neural network model, which consists of two parts: an encoder and a decoder, and can learn the features of the input data; ultimately, it can obtain the binding protein features of each drug. Step 3: Adaptive Learning 3.1) The drug molecule structure information features and drug-bound protein features are concatenated and fed into the adaptive learning module. In the adaptive learning module, N+2 adjacency matrices are first formed, where N represents the number of adjacency matrices corresponding to each adverse reaction type. Each adjacency matrix is ​​formed by connecting drug pairs with interactions in an adverse reaction event. The two adjacency matrices are represented by the drug molecule structure information similarity feature map and the drug-bound protein similarity feature map. The similarity feature map is obtained through metric learning, with the following formula: ， ， in Represents an adjacency matrix. and Each corresponds to a drug in the adjacency matrix and medicine , It is a cosine similarity function with K weight parameters. and Corresponding drugs and medicine Features For the threshold, These are learnable weight parameters; thus, drug molecule structure information similarity feature maps and drug-bound protein similarity feature maps are constructed by using the similarity of drug molecule structure information features and the similarity of drug-bound protein features. 3.2) The obtained N+2 adjacency matrices are subjected to adaptive transformation. The adaptive transformation is achieved by adaptively learning the most effective meta-paths associated with drug-adverse drug reactions: firstly, a learnable weight matrix is ​​used. Multiplying and summing the results with N+2 adjacency matrices yields a subgraph, where weight parameters are learned from each of the N+2 adjacency matrices. ,in The weight parameters are 1x1. A larger value indicates that the corresponding adjacency matrix is ​​more important to the prediction result, thus integrating all important adjacency matrix information into a single subgraph. The formula is: ； 3.3) Step 3.2) is a single-layer adaptive transformation, that is, the subgraph formed by connecting neighboring nodes with a distance of 1 hop. Layer adaptive transformation and matrix multiplication yield a length of The metapath, that is, for nodes The subgraph formed by connecting neighboring nodes with hop distance is defined by the following formula: ， This adaptive learning reaches a state with The skip distance subgraph can integrate effective drug-adverse reaction type correlation information, where... It is a constant; 3.4) Multi-channel learning is employed, i.e., num_channels are set for learning, ultimately learning num_channels with... The subgraph of jump distance; 3.5) In num_channels with The graph convolutional neural network is used to train and learn on the subgraph of the jump distance. It will fuse the neighbor information of the nodes in each channel subgraph to form the final node feature matrix, and then stack the node feature matrices of multiple channels to form the final feature matrix of the drug. Step 4: Multilayer Perceptron Prediction 4.1) Extract features of drug pairs that will cause adverse drug reactions from the final feature matrix. and Furthermore, the features of the two drugs are concatenated and then fed into a perceptron consisting of two fully connected neural networks for learning, using the following formula: ， ， in and Represents the learnable weight parameters. and It is a bias vector. It is the Sigmoid activation function. This is the final predicted value. The model uses cross-entropy as the loss function for training and learning.

Images