Adverse drug reaction prediction method based on message passing model

Abstract

The present invention discloses a method for predicting adverse drug reactions based on a message passing model, which relates to the field of adverse drug reaction prediction. The method includes encoding drug data labels into graph data and improving a new message aggregation mechanism based on the message passing model to improve message aggregation accuracy and shorten agent operation time. Simultaneously, the stability of the agent is improved by improving the agent structure. This method addresses the shortcomings of existing prediction methods in terms of drug data label processing, agent accuracy, operation speed, and stability. This method can effectively improve the accuracy of adverse drug reaction predictions and the speed and stability of agent operation.

Description

Technical Field

[0001] The present invention belongs to the field of adverse drug reaction prediction, and in particular relates to a method for predicting adverse drug reactions based on a message passing model. Background Art

[0002] According to the World Health Organization's Collaborating Centre for International Drug Monitoring, adverse drug reactions are defined as harmful reactions unrelated to the intended use of a drug when taken at normal doses. These drugs are primarily used to prevent, diagnose, treat disease, or regulate physiological functions.

[0003] With the in-depth study and continuous advancement of adverse drug reactions, a large number of computational agents have been proposed. However, most of these methods only perform simple calculations on the nodes and neighboring nodes of the drug structure data in the drug dataset, and do not fully utilize the drug label information. This leads to low accuracy, slow operation speed, and poor stability of current methods.

[0004] Therefore, there is an urgent need for a method for predicting adverse drug reactions that solves the above technical problems. Summary of the Invention

[0005] The technical problem to be solved by the present invention is to provide a drug adverse reaction prediction method based on a message passing model, which solves the problems of low accuracy, slow operation speed and poor stability of drug adverse reaction prediction methods in the prior art.

[0006] The present invention adopts the following technical solutions to solve the above technical problems:

[0007] The method for predicting adverse drug reactions based on the message passing model includes the following steps:

[0008] Step 1: Convert the chemical molecular structure feature information of the drugs in the drug dataset into graph data, and then convert this graph data into a dataset that can be calculated by the intelligent agent;

[0009] Step 2: Pass the dataset to the agent;

[0010] Step 3: Extract the drug feature data from the dataset in the agent, build a new message aggregation mechanism based on the message passing model, obtain dynamic update node parameters, and weight and filter the specific drug feature data;

[0011] Step 4: Input the screened drug characterization information into the residual network structure for updating to obtain new drug characterization;

[0012] Step 5: Put the updated drug information data into the scoring function for scoring, and output the predicted adverse drug reaction results;

[0013] Step 6: Take the adverse drug reaction results as input, calculate using the cross entropy function, and update the agent.

[0014] The intelligent agent is a prediction intelligent agent including a fully connected network, a message passing model, a scoring function, and a cross entropy function.

[0015] In step 2, the dynamically updated node parameter π is used to weight and filter the drug characterization information in the graph data. The formula is as follows:

[0016]

[0017]

[0018]

[0019] Where:

[0020] k is the number of rows in the matrix after linear separation; i and j are the target node and its neighbor nodes in the drug graph data respectively;

[0021] (l) is the agent after the l-1th iteration; g(·) is a learnable linear function;

[0022] These are two trainable matrices separated from the linear layer. The dimensions of the matrices are determined by the dimensions of the feature matrix and the label matrix in different datasets.

[0023] z γ ,z β is the row value in the γ and β matrices, D γ ,D β is the dimension of matrix γ, β matrix, and x∈D γ ,y∈D β ;

[0024] f(·) is the maximum value function in the matrix, are the maximum values ​​in each row of the drug characteristic information data matrix obtained after f(·).

[0025] In step 3, the drug characterization information after screening is input into the agent structure with the residual agent for updating, and the formula is expressed as:

[0026]

[0027] Where: is the representation of the target node i after l iterations; W is the learnable weight, σ1, σ2 are the activation functions PReLU, They are respectively the dynamically updated parameters obtained after screening the drug characteristic information data matrix.

[0028] In step 4, the updated drug pair representation is put into the scoring function S for scoring, which is expressed as follows:

[0029] S(d m ,d n ,r)=σ(exp(W mn (h m ||h n ))⊙u r );

[0030] Where: d m ,d n Represent the two drugs that need to be predicted; W mn is a trainable weight matrix; h m ,h n is the two target prediction drug characteristic information obtained after the agent training update; r is the relationship type for predicting adverse drug reactions; u r is the high-dimensional vector embedding of r; σ is the activation function Sigmoid; ‖ is the connection operation.

[0031] In step 5, the cross entropy function is used to calculate and update the agent, and its formula is expressed as:

[0032]

[0033] Where: L represents the function of minimizing cross entropy loss. It is known that |μ represents the number of rows of the label matrix of whether there is an antagonistic reaction between drugs m and n, and y i Indicates the label of a known antagonistic reaction, when y i =1, indicating drug d m and d n There is an interaction relationship between them; vice versa, that is, when y i When it is not equal to 1, it means drug d m with d n There is no interaction between them.

[0034] A computer storage medium stores computer instructions, which are used to execute all or part of the steps of the method when called.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] 1. A new message aggregation method is used. The affine function between the target node and its neighboring nodes is used as the attention coefficient, and the vector representations of the neighboring nodes are weighted summed. This allows for more flexible adjustment of the contribution of different neighboring nodes to the target node, improving the efficiency and accuracy of message aggregation.

[0037] 2. We propose a new feature-encoded message passing architecture that concatenates the target node representation with the representations of its neighboring nodes and generates an encoding vector that controls the target node's features through a linear layer. This architecture makes better use of the information between the target node and its neighboring nodes, generating an affine function that is more suitable for message passing.

[0038] 3. In the process of message transmission, the residual idea is also added to alleviate the smoothing phenomenon of intelligent agent nodes, thereby improving the stability of the intelligent agent. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 This is a flow chart of the prediction method of the present invention. DETAILED DESCRIPTION

[0040] In order to enable people in this technical field to better understand the solution of this application, the technical solution in the embodiment of this application will be clearly and completely described below in combination with the drawings in the embodiment of this application. Obviously, the described embodiment is only an embodiment of a part of this application, not all embodiments.

[0041] Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making any creative work should fall within the scope of protection of this application.

[0042] The purpose of the present invention is to provide a method for predicting adverse drug reactions based on a message passing model to solve the problem that existing prediction methods cannot fully utilize drug label information, resulting in low accuracy, slow operation speed and poor stability.

[0043] The method for predicting adverse drug reactions based on the message passing model includes the following steps:

[0044] Step 1: Convert drug labels into graph data and pass all graph data to the agent;

[0045] Step 2: Extract drug representation information from graph data in the agent, build a new message aggregation mechanism based on the message passing model, obtain dynamically updated node parameters, and weight and filter the drug representation information of the graph data;

[0046] Step 3: Input the screened drug characterization information into the residual network structure for updating to obtain new drug characterization;

[0047] Step 4: Put the updated drug representation into the scoring function for scoring, and output the predicted adverse drug reaction results;

[0048] Step 5: Take the adverse drug reaction results as input, calculate using the cross entropy function, and update the agent.

[0049] Specific embodiments, such as Figure 1 As shown,

[0050] A method for predicting adverse drug reactions based on a message passing model first obtains all drug information datasets as training data, and then performs the following steps:

[0051] S01: Use the data class in the PyG library to convert the chemical molecular structure feature information of drugs in the SMILES-encoded drug dataset into graph data through the rdkit tool, and then use the PYG tool to convert these graph data into datasets that can be calculated by the intelligent agent, and then pass these graph data to the intelligent agent to facilitate subsequent model application; the intelligent agent includes a prediction intelligent agent composed of a fully connected network, a message passing model, a scoring function, and a cross-entropy function.

[0052] S02: Apply a fully connected layer in the agent to extract drug feature information from the graph data in step S01. Initialize linear variation parameters through the linear layer to filter drug features. The filtered drug features are sent to the message passing model. Based on the message passing model, a new message aggregation mechanism is improved to obtain the dynamically updated node parameter π to weight and filter the drug representation information of the graph data.

[0053] The linear variation parameter, described as adding a linear layer, can also be understood as a randomly initialized matrix whose dimensions match those of the drug dataset. This matrix is ​​then split into two parameters, alpha and beta, which are dynamically updated parameters. These two dynamically updated parameters weight the content of the drug dataset. Specifically in code, the linear update is the opposite of the nonlinear update. Those skilled in the art understand that this linear variation parameter can also be considered a neural network hyperparameter. The only requirement is that its dimensions must match those of the drug dataset to ensure that the input and output correspond when the matrices are multiplied.

[0054] The weighting is a dynamically updated weighting, which is mainly based on the drug feature information in the dataset, and only the key features are weighted. The dynamically updated node parameter π is used to weight and filter the drug representation information in the graph data. The screening criteria are as follows:

[0055]

[0056]

[0057] Where, is the feature information of drug i in the drug graph data after l rounds of updates; Represents the first row of data in the drug i graph data matrix, It is a dynamically updated parameter;

[0058] The weighted formula is expressed as:

[0059]

[0060]

[0061]

[0062] Where:

[0063] k is the number of rows in the matrix after linear separation; i and j are the predicted drugs in the drug map data and the drugs currently known to have antagonistic reactions with the drug, respectively;

[0064] (l) is the agent after the l-1th iteration; g(·) is a learnable linear function;

[0065] They are two trainable matrices separated from the linear layer. The dimensions of the matrices are determined by the dimensions of the feature matrix and the label matrix in different datasets; θ k are the hyperparameters in the neural network, specifically the dimensions of the input layer, hidden layer, and output layer.

[0066] z γ ,z β The γ and β matrices are obtained by looping functions from top to bottom according to each row of data, D γ ,D β is the dimension of matrix γ, β matrix, and x∈D γ ,y∈D β ;z x ,z y Represents z γ ,z β Maximum value in the row;

[0067] f(·) is the function for finding the maximum value in the matrix.

[0068] are the maximum values ​​in each row of the drug characteristic information data matrix obtained after f(·).

[0069] S03: Input the screened drug characterization information into the agent structure with residual agent for updating. The formula is expressed as:

[0070]

[0071] Where: is the representation of the target node i after l iterations; W is the learnable weight used to obtain new drug representations, σ1, σ2 are the activation functions PReLU; N i is the set of neighbor nodes j of node i, W k is the learnability weight (i.e., hyperparameter) of the corresponding neighbor node j.

[0072] S04: Put the updated drug pair representation into the scoring function S for scoring, and output the predicted adverse drug reaction result; the formula of the scoring function is expressed as:

[0073] S(d m ,d n ,r)=σ(W mn (h m ||h n )⊙u r );

[0074] Where: d m ,d n Represent the two drugs that need to be predicted; W mn is a trainable weight matrix; h m ,h n is the two target prediction drug characteristic information obtained after the agent training update; r is the relationship type for predicting adverse drug reactions; u r is the high-dimensional vector embedding of r; σ is the Sigmoid activation function; ‖ is the concatenation operation. If the predicted value of S is greater than 0.5, it is considered that there is an adverse reaction between the two drugs.

[0075] S05: Finally, the agent is updated by taking the adverse drug reaction results as input and calculating using the cross entropy function; the formula is expressed as:

[0076]

[0077] Where: L represents the function of minimizing cross entropy loss. It is known that |μ represents the number of rows of the label matrix of whether there is an antagonistic reaction between drugs m and n, and y i Indicates the label of a known antagonistic reaction, when y i =1, indicating drug d m and d n There is an interaction relationship between them; vice versa, that is, when y i When it is not equal to 1, it means drug d m with d n There is no interaction between them.

[0078] This not only improves the accuracy and stability of the intelligent agent in predicting adverse drug reactions, but also greatly shortens the running time.

[0079] In order to further understand this solution, two specific examples are used for illustration.

[0080] This study was based on the DrugBank dataset. The dataset was first split into two independent subsets: DrugBank01 and DrugBank02, ensuring mutual exclusivity of drug combinations between the subsets. The model was first trained using the DrugBank011 dataset. Then, drug data were randomly sampled from the DrugBank02 dataset for computation, and the model generated results. Finally, the model was validated using data from the DrugBank02 dataset.

[0081] The data were processed using the method in this application, and the model was optimized using the DrugBank01 dataset as the training set.

[0082] After the model training is completed, it is calculated:

[0083] 1. The SMILES encoding of DB00608 is

[0084] The SMILES encoding of CCN(CC)CCCC(C)NC1=CC=NC2=CC(Cl)=CC=C12 and DB00358 is

[0085] OC(C1CCCCN1)C1=CC(=NC2=C1C=CC=C2C(F)(F)F)C(F)(F)F may cause drug interactions.

[0086] The DrugBank02 dataset confirmed that taking the two compounds simultaneously would cause adverse reactions.

[0087] 2. The SMILES code of DB01059 is

[0088] CCN1C=C(C(O)=O)C(=O)C2=CC(F=C(C=C12)N1CCNCC1 and DB01032 have SMILES encoding:

[0089] CCCN(CCC)S(=O)(=O)C1=CC=C(C=C1)C(O)=O may cause drug interactions.

[0090] The DrugBank02 dataset confirmed that taking the two compounds simultaneously would increase serum concentrations.

[0091] Compared with the prior art, the present invention has the following significant advantages:

[0092] 1) A new message aggregation method is adopted, which uses the affine function between the target node and its neighbor nodes as the attention coefficient and performs a weighted summation on the vector representations of the neighbor nodes. This allows for more flexible adjustment of the contribution of different neighbor nodes to the target node, improving the efficiency and accuracy of message aggregation.

[0093] 2) A new feature-encoded message-passing architecture is proposed. This concatenates the representation of the target node with that of its neighboring nodes and generates an encoding vector that controls the target node's features through a linear layer. This architecture makes better use of the information between the target node and its neighboring nodes, generating an affine function that is more suitable for message passing. Furthermore, a residual concept is incorporated into the message passing process to mitigate agent node smoothing, thereby improving the agent's stability.

[0094] Those skilled in the art should understand that they can implement variations by combining the prior art and the above embodiments. Such variations do not affect the essence of this solution and are not described in detail here.

[0095] It should be understood that this solution is not limited to the specific implementation methods described above. Devices and structures not described in detail should be understood to be implemented in a common manner in the art. Any person skilled in the art can, without departing from the scope of this solution, use the methods and technical content disclosed above to make many possible changes and modifications to this solution, or modify it into equivalent embodiments with equivalent changes, without affecting the essence of this solution. Therefore, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of this solution without departing from the content of this solution are still within the scope of protection of this solution.

Claims

1. A method for predicting adverse drug reactions based on a message passing model, characterized by: The steps include: Step 1: Convert the chemical molecular structure feature information of the drugs in the drug dataset into graph data, and then convert this graph data into a dataset that can be calculated by the intelligent agent; Step 2: Pass the dataset to the agent; Step 3: Extract the drug feature data in the dataset in the agent, and build a new message aggregation mechanism based on the message passing model to obtain dynamic update node parameters, and weight and filter the specific drug feature data; use the dynamic update node parameters to The drug characterization information in the graph data is weighted and screened, and the screening criteria are as follows: ； ； Where, For drugs in the drug graph data The characteristic information of Feature information after round update; Indicates drug The first row of data in the graph data matrix, It is a dynamically updated parameter; The weighted formula is as follows: ； ； ； Where: is the number of rows of the matrix after linear separation; are the target node and its neighbor nodes in the drug graph data respectively; The agent has gone through the iterations; is a learnable linear function; These are two trainable matrices separated from the linear layer. The dimensions of the matrices are determined by the dimensions of the feature matrix and the label matrix in different datasets. for The values ​​in the matrix are taken by row, is a matrix The dimensions of the matrix, and ; To find the maximum value function in a matrix, , After The maximum value in each row of the drug characteristic information data matrix obtained later; Step 4: Input the screened drug characterization information into the residual network structure for updating to obtain new drug characterization; Step 5: Put the updated drug information data into the scoring function for scoring, and output the predicted adverse drug reaction results. The formula is expressed as: ； Where: Represent the two drugs that need to be predicted; is a trainable weight matrix; Predict drug property information for the two targets finally updated after training the agent; Types of relationships for predicting adverse drug reactions; yes High-dimensional vector embedding of is the activation function Sigmoid; It is a join operation; Step 6: Take the adverse drug reaction results as input, calculate using the cross entropy function, and update the agent.

2. The method for predicting adverse drug reactions based on a message passing model according to claim 1, wherein: The intelligent agent is a prediction intelligent agent including a fully connected network, a message passing model, a scoring function, and a cross entropy function.

3. The method for predicting adverse drug reactions based on a message passing model according to claim 1, wherein: In step 4, the drug characterization information after screening is input into the agent structure with the residual agent for updating, and the formula is expressed as: ； Where: Target node In passing Representation after iterations; are learnable weights, , is the activation function PReLU, , They are respectively the dynamically updated parameters obtained after screening the drug characteristic information data matrix.

4. The method for predicting adverse drug reactions based on a message passing model according to claim 1, wherein: In step 6, the cross entropy function is used to calculate and update the agent, and its formula is expressed as: ； Where: Represents minimizing the cross entropy loss function, currently known Indicates drug and The number of rows in the label matrix indicating whether there is an antagonistic reaction between them, Indicates a label for a known antagonistic reaction, when When, it means medicine and There is an interactive relationship between When it is not equal to 1, it means drug and There is no interaction between them.

5. A computer storage medium, characterized in that: The computer storage medium stores computer instructions, which are used to execute all or part of the steps of the method according to any one of claims 1 to 4 when called.

Images