Drug-target interaction mechanism prediction method and system based on ternary comparison professional knowledge decision

Abstract

The invention discloses a drug-target interaction mechanism prediction method and system based on ternary comparison professional knowledge decision, and the method comprises the steps: firstly, carrying out the preprocessing of initial drug target action mechanism data, obtaining the complete information of drug molecules and target proteins, and screening a head class and a tail class; secondly, an original multi-classification long tail problem is decomposed into a plurality of simple subtasks, each subtask comprises two original categories and a designed universal category # imgabs0 #, a training strategy is provided, two different encoders are used for encoding drug molecules and target proteins respectively, and respective embedded vectors are obtained; constructing a prediction model by adopting a plurality of linear layers, taking the obtained drug and protein embedding vectors as the input of the prediction model, carrying out adaptive learning rate optimization algorithm optimization model parameters by using a cross entropy classification loss function for a prediction category probability score, and proposing a decision voting module based on category balance; and finally, obtaining a drug-target interaction mechanism prediction model through training, and performing an interaction mechanism prediction task on a drug-target pair in a real scene by using the model. The method can effectively solve the problem of predicting the action mechanism of the drug target in long-tail distribution.

Description

Technical Field

[0001] The present invention belongs to the field of bioinformatics and long-tail learning, and relates to a drug-target interaction mechanism prediction method and system based on ternary comparative professional knowledge decision-making. Background Art

[0002] Drug development is a complicated and time-consuming project. In order to develop reliable and safe drugs, researchers need to carry out a lot of experimental verification, which consumes a lot of manpower, material resources and financial resources. According to statistics, it takes an average of about 15 years and about 2 billion US dollars to manufacture a small molecule drug. As a key step in drug discovery and repositioning, drug-target interaction prediction can find new targets for existing drugs and find new therapeutic uses for drugs. In the initial stage of drug development, if it can be correctly predicted that there is an interaction between the drug and certain targets in the body, it can be proved that the drug can act on these targets after entering the body. Researchers only need to focus on specific targets and avoid aimlessly screening from a large number of candidates. Therefore, accurate and rapid identification of drug-target interactions is conducive to shortening the drug development cycle and reducing R&D costs, especially in the response to sudden diseases, it can discover new clinical uses of drugs that have been rigorously tested and verified, which has important economic and social value.

[0003] Existing related algorithms can quickly and accurately predict the interaction between drugs and specific targets, that is, whether the drug molecules can successfully bind to certain targets in the human body. However, drugs do not work simply by binding to targets, but affect the subsequent biological activities of target receptors from different mechanisms of action (excitation, antagonism, etc.) and to different degrees (complete, partial, etc.). Therefore, the interaction between drugs and targets should be studied in depth as a more complex multi-category mechanism prediction problem, but there are currently almost no deep learning models / algorithms for predictive analysis of all known mechanisms of action. In addition, since there are too few existing mechanism data and the entire data set presents a long-tail distribution, the prediction performance of existing algorithms in the tail category is very challenging. The existing technology has the following disadvantages:

[0004] 1) Insufficient research on mechanism prediction: Studying the binding of drugs to human targets is essentially studying the various mechanisms of action produced by drug molecules acting on targets. Most existing methods only focus on whether the two can bind. However, it is necessary to accurately determine the specific mechanism of action between the two (such as excitation, antagonism, etc.) in order to scientifically and accurately accelerate the drug repositioning process. Therefore, how to accurately determine "what mechanism of action the drug has on the target" has brought great challenges to the research related to drugs and human targets.

[0005] 2) Inaccurate prediction of long-tail distribution data: Some drug-target mechanisms of action (such as allosteric regulation) are relatively rare, so the amount of relevant data is extremely scarce compared to other mainstream action categories (such as activation, antagonism, etc.). Therefore, it is difficult to complete accurate mechanism prediction based on small sample data in the case of class imbalance. Summary of the invention

[0006] To solve the above problems, the present invention provides a method and system for predicting drug-target interaction mechanism based on ternary comparison professional knowledge decision-making.

[0007] The present invention takes into account the difficulty of learning long-tail data, models the original multi-classification task into multiple simple sub-tasks based on ternary comparison, proposes a sub-task training strategy based on ternary comparison professional knowledge decision, and adopts a decision voting mechanism based on category balance to obtain the final multi-classification prediction result on the basis of obtaining all sub-task prediction results, so as to be applied to the drug-target interaction mechanism prediction task. The first aspect of the invention relates to:

[0008] The drug-target interaction mechanism prediction method and system based on ternary comparison professional knowledge decision-making includes the following steps:

[0009] 1. Preprocessing of drug-target action mechanism dataset;

[0010] Based on the existing drug-target interaction data set, further collect and organize samples related to various mechanisms of action such as agonism and antagonism. After retrieving all qualified samples, the raw data must be preprocessed. The specific preprocessing steps include:

[0011] Step 11: Only drug-target pairs that are relevant to humans and have complete field information are retained.

[0012] Step 12: Use bioinformatics tools such as RDKit to determine whether the SMILES of the drug molecule is illegal.

[0013] Step 13: For proteins for which sequence information does not exist, the sequence of the target protein is searched across platforms in the UniProt database using a unique identifier.

[0014] Step 14: Analyze the mechanism type field of the processed data set (such as including agonists and inhibitors), and filter the head class and tail class according to the data volume.

[0015] Finally, a large-scale data set with "drug molecule-target protein sequence-action mechanism type" as the basic data elements was obtained, of which the first two were used as input to the model and the latter was used as the true label for the predicted output.

[0016] 2. Professional knowledge model training based on ternary comparison;

[0017] Step 21: Subtask data division. The original drug target mechanism prediction task is labeled as a multi-classification task with C categories. Each subtask aims to classify only two categories, resulting in a total of C*(C-1) / 2 subtasks. Each subtask is trained by a specific expertise model to extract knowledge related to the corresponding two categories (category 1, category 2). In order to ensure the accuracy of the expertise model prediction, an additional category is introduced. It is used to represent samples that do not belong to the two selected categories. Therefore, samples related to the two categories and samples belonging to the categories are extracted from the overall data set. samples to achieve the prediction task based on ternary comparison. Finally, the training dataset of this subtask is the combination of the above three datasets.

[0018] Step 22: Expertise model training. The expertise model is responsible for performing a simple three-way comparison task to determine whether the mechanism of a drug-target pair belongs to category 1, category 2, or Each expertise model is based on a classic encoder-decoder architecture. The encoder module includes two encoders that process drug molecules and target proteins respectively, thereby obtaining representation vectors of drugs and proteins:

[0019] 1) For drug molecule m, it is usually represented as a molecular graph, and then the naive GCN model is used to iteratively update its initial biochemical feature representation Specifically, the node embedding of layer l is Updated to:

[0020]

[0021] in is the output node embedding of the (l-1)th layer; is the trainable weight of the lth layer; A is the adjacency matrix of the molecular graph; σ(·) represents the nonlinear activation layer. After L iterative calculations, the final node embedding vector can be obtained Finally, a global pooling layer is applied to obtain the representation of the entire drug molecule graph.

[0022] 2) For the target protein t, it is usually represented as an amino acid sequence of unequal length, and then a 1D CNN model is used to iteratively update its initial biochemical feature representation Specifically, the node embedding of layer l is Updated to:

[0023]

[0024] in is the output embedding of the (l-1)th layer; k (l) Represent the number of input channels, output channels and convolution kernel size of the lth layer respectively; σ(·) represents the nonlinear activation layer. After K iterations, the final protein representation can be obtained.

[0025] The decoder module takes the representation vector of the combination of drugs and proteins from the encoder as input and uses multiple linear layers to obtain a 3D output, where each dimension represents 1, 2, or Class, the probability score of the node is obtained through the Softmax nonlinear layer The prediction of the category probability score is optimized using the cross entropy classification loss function. The specific formula is as follows:

[0026]

[0027] Where y is the true label of the sample. If the interaction between the samples is of type 1, the value is 0; if the interaction is of type 2, the value is 1; if the interaction between the samples is neither of type 1 nor of type 2, the value is -1.

[0028] 3. Decision voting based on category balance;

[0029] After obtaining the prediction labels of the subtasks from all the expertise models, a decision voting strategy based on class balance is adopted to effectively integrate the prediction results of each expertise model. Specifically, C*(C-1) / 2 initial prediction results Q are obtained from the expertise models, which are defined as follows:

[0030] Q=(q 12 ,q 13 ,…,q 1c ,q 23 ,…,q 2C ,…,q C-1,C ), (4)

[0031] where q i,j represents the output of the ternary comparison task between the i-th and j-th categories, and q i,j ∈{-1,0,1}. Next, based on the three possible outputs, update the final voting vector Y∈R for C categories C As follows: 1) If q i,j is 0, indicating that category i is the output label, and the reward will be assigned to the voting score of category i, denoted as Y i ; 2) If q i,j is 1, indicating that category j is the output label, and the reward will be assigned to the voting score of category j, denoted as Y j ; 3) If qi,j -1, indicating category is the output label, and the penalty will be assigned to Y i and Y j .

[0032] Finally, after traversing the initial predictions of all professional models, the voting scores of all categories are counted. Due to the imbalance between categories, a class balance weight vector H∈R based on the number of samples is used. C To calculate the final score, the category with the most votes is selected as the final prediction result.

[0033] 4. Achieve the task of predicting drug-target interaction mechanisms;

[0034] All samples in the drug-target interaction mechanism dataset are divided into training set, validation set and test set according to the ratio of 80%, 10% and 10%. The data is divided into batches and sent to the model for multiple epoch training. The adaptive learning rate optimization algorithm is used to iteratively update the model parameters. In this process, the performance of each epoch model is recorded, and the performance of the model is evaluated by using the performance of the accuracy and F1 value. The model parameters with the best average performance are selected to finally obtain the drug-target interaction mechanism prediction model.

[0035] The second aspect of the present invention relates to a drug-target interaction mechanism prediction system based on ternary comparison expertise decision: it includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, they are used to implement the drug-target interaction mechanism prediction method based on ternary comparison expertise decision of the present invention.

[0036] The third aspect of the present invention relates to a computer-readable storage medium having a program stored thereon, which, when executed by a processor, implements the drug-target interaction mechanism prediction method based on ternary comparison expertise decision of the present invention.

[0037] The advantages of the present invention are: decomposing the original multi-classification problem into multiple simple subtasks, and effectively solving the problem of drug target mechanism prediction with long-tail distribution through a novel professional knowledge decision-making method based on ternary comparison. Secondly, a professional knowledge training strategy based on ternary comparison is designed: first, a divide-and-conquer strategy is adopted to decompose the original task into subtasks, ensuring that the main category does not occupy the resources of the tail category and reducing the difficulty of the original task; at the same time, the category is introduced The decision boundaries for the subtasks of categories 1 and 2 are determined and supplemented by new samples to generate more robust representations for the tail categories. Then, a decision voting module based on category balance is proposed to further integrate professional knowledge. Finally, a flexible model evaluation method is adopted to comprehensively evaluate the performance of the model by taking into account multiple indicators such as accuracy and F1 value. Through this mechanism prediction model, the ability to distinguish similar mechanism types can be enhanced, and a more effective and robust feature representation can be provided for the tail categories, thereby accurately and efficiently completing the task of predicting the mechanism of action of drug targets. In addition, the detailed exploration of the evolvability and generalization of this method promotes its practicality and effectiveness in actual scenarios, thereby providing an effective computer-aided tool for drug design, achieving the effects of shortening the drug development cycle and reducing R&D costs. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 is a flow chart of the method of the present invention.

[0039] Figure 2 It is a specific professional knowledge model architecture diagram of the present invention. DETAILED DESCRIPTION

[0040] The technical solution of the present invention is clearly and completely explained and described below.

[0041] Example 1

[0042] This embodiment relates to a method for predicting drug-target interaction mechanism based on ternary comparison professional knowledge decision-making, comprising the following steps:

[0043] 1. Preprocessing of drug-target action mechanism dataset;

[0044] Based on the existing drug-target interaction data set, further collect and organize samples related to various mechanisms of action such as agonism and antagonism. After retrieving all qualified samples, the raw data must be preprocessed. The specific preprocessing steps include:

[0045] Step 11: Only drug-target pairs that are relevant to humans and have complete field information are retained.

[0046] Step 12: Use bioinformatics tools such as RDKit to determine whether the SMILES of the drug molecule is illegal.

[0047] Step 13: For proteins for which sequence information does not exist, the sequence of the target protein is searched across platforms in the UniProt database using a unique identifier.

[0048] Step 14: Analyze the mechanism type field of the processed data set (such as including agonists and inhibitors), and filter the head class and tail class according to the data volume.

[0049] Finally, a large-scale data set with "drug molecule-target protein sequence-action mechanism type" as the basic data elements was obtained, of which the first two were used as input to the model and the latter was used as the true label for the predicted output.

[0050] 2. Professional knowledge model training based on ternary comparison;

[0051] Step 21: Subtask data division. The original drug target mechanism prediction task is labeled as a multi-classification task with C categories. Each subtask aims to classify only two categories, resulting in a total of C*(C-1) / 2 subtasks. Each subtask is trained by a specific expertise model to extract knowledge related to the corresponding two categories (category 1, category 2). In order to ensure the accuracy of the expertise model prediction, an additional category is introduced. It is used to represent samples that do not belong to the two selected categories. Therefore, samples related to the two categories and samples belonging to the categories are extracted from the overall data set. samples to achieve the prediction task based on ternary comparison. Finally, the training dataset of this subtask is the combination of the above three datasets.

[0052] Step 22: Expertise model training. The expertise model is responsible for performing a simple three-way comparison task to determine whether the mechanism of a drug-target pair belongs to category 1, category 2, or Each expertise model is based on a classic encoder-decoder architecture. The encoder module includes two encoders that process drug molecules and target proteins respectively, thereby obtaining representation vectors of drugs and proteins:

[0053] 1) For drug molecule m, it is usually represented as a molecular graph, and then the naive GCN model is used to iteratively update its initial biochemical feature representation Specifically, the node embedding of layer l is Updated to:

[0054]

[0055] in is the output node embedding of the (l-1)th layer; is the trainable weight of the lth layer; A is the adjacency matrix of the molecular graph; σ(·) represents the nonlinear activation layer. After L iterative calculations, the final node embedding vector can be obtained Finally, a global pooling layer is applied to obtain the representation of the entire drug molecule graph.

[0056] 2) For the target protein t, it is usually represented as an amino acid sequence of unequal length, and then a 1D CNN model is used to iteratively update its initial biochemical feature representation Specifically, the node embedding of layer l is Updated to:

[0057]

[0058] in is the output embedding of the (l-1)th layer; k (l) Represent the number of input channels, output channels and convolution kernel size of the lth layer respectively; σ(·) represents the nonlinear activation layer. After K iterations, the final protein representation can be obtained.

[0059] The decoder module takes the representation vector of the combination of drugs and proteins from the encoder as input and uses multiple linear layers to obtain a 3D output, where each dimension represents 1, 2, or Class, the probability score of the node is obtained through the Softmax nonlinear layer The prediction of the category probability score is optimized using the cross entropy classification loss function. The specific formula is as follows:

[0060]

[0061] Where y is the true label of the sample. If the interaction between the samples is of type 1, the value is 0; if the interaction is of type 2, the value is 1; if the interaction between the samples is neither of type 1 nor of type 2, the value is -1.

[0062] 3. Decision voting based on category balance;

[0063] After obtaining the prediction labels of the subtasks from all the expertise models, a decision voting strategy based on class balance is adopted to effectively integrate the prediction results of each expertise model. Specifically, C*(C-1) / 2 initial prediction results Q are obtained from the expertise models, which are defined as follows:

[0064] Q=(q 12 ,q 13 ,…,q 1C ,q 23 ,…,q 2C ,…,q C-1,C ), (4)

[0065] where q i,j represents the output of the ternary comparison task between the i-th and j-th categories, and q i,j ∈{-1,0,1}. Next, based on the three possible outputs, update the final voting vector Y∈R for C categories C As follows: 1) If q i,jis 0, indicating that category i is the output label, and the reward will be assigned to the voting score of category i, denoted as Y i ; 2) If q i,j is 1, indicating that category j is the output label, and the reward will be assigned to the voting score of category j, denoted as Y j ; 3) If q i,j -1, indicating category is the output label, and the penalty will be assigned to Y i and Y j .

[0066] Finally, after traversing the initial predictions of all professional models, the voting scores of all categories are counted. Due to the imbalance between categories, a class balance weight vector H∈R based on the number of samples is used. C To calculate the final score, the category with the most votes is selected as the final prediction result.

[0067] 4. Achieve the task of predicting drug-target interaction mechanisms;

[0068] All samples in the drug-target interaction mechanism dataset are divided into training set, validation set and test set according to the ratio of 80%, 10% and 10%. The data is divided into batches and sent to the model for multiple epoch training. The adaptive learning rate optimization algorithm is used to iteratively update the model parameters. In this process, the performance of each epoch model is recorded, and the performance of the model is evaluated by using the performance of the accuracy and F1 value. The model parameters with the best average performance are selected to finally obtain the drug-target interaction mechanism prediction model.

[0069] Example 2

[0070] The present embodiment relates to a drug-target interaction mechanism prediction system based on ternary comparison professional knowledge decision: it includes a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the drug-target interaction mechanism prediction method based on ternary comparison professional knowledge decision of Example 1.

[0071] Example 3

[0072] This embodiment relates to a computer-readable storage medium having a program stored thereon. When the program is executed by a processor, the drug-target interaction mechanism prediction method based on ternary comparison professional knowledge decision of Embodiment 1 is implemented.

[0073] The contents described in the embodiments of this specification are merely an enumeration of implementation forms of the inventive concept. The protection scope of the present invention should not be regarded as limited to the specific forms described in the embodiments. The protection scope of the present invention also extends to equivalent technical means that can be conceived by those skilled in the art based on the inventive concept.

Claims

1. A drug-target interaction mechanism prediction method based on ternary comparison professional knowledge decision-making includes the following steps: S1. Preprocess the drug-target action mechanism dataset: Based on the existing drug-target interaction dataset, collect and organize samples related to various action mechanisms of agonism and antagonism; after retrieving all qualified samples, preprocess the raw data; S2. Train an expertise model based on ternary comparison; S3. Class-balanced decision voting: After obtaining the prediction labels of the subtasks from all the expertise models, a class-balanced decision voting strategy is adopted to effectively integrate the prediction results of each expertise model. S4. Implement the task of drug-target interaction mechanism prediction.

2. The method for predicting drug-target interaction mechanism based on ternary comparison professional knowledge decision-making according to claim 1, characterized in that: Step S1 specifically includes: S11: Only drug-target pairs that are relevant to humans and have complete field information are retained; S12: Use RDKit and other bioinformatics tools to determine whether the SMILES of drug molecules are illegal; S13: For protein sequence information that does not exist, the sequence of the target protein is searched across platforms in the UniProt database using a unique identifier; S14: Analyze the mechanism type field of the processed data set and filter the head and tail classes according to the data volume; Finally, a large-scale data set with "drug molecule-target protein sequence-action mechanism type" as the basic data elements was obtained, of which the first two were used as input to the model and the latter was used as the true label for the predicted output.

3. The method for predicting drug-target interaction mechanism based on ternary comparison professional knowledge decision-making as claimed in claim 1, characterized in that: Step S2 specifically includes: S21: Subtask data partitioning: The original drug target mechanism prediction task is labeled as a multi-classification task with C categories; each subtask aims to classify only two categories, resulting in a total of C*(C-1) / 2 subtasks; each subtask is trained by a specific expertise model to extract knowledge related to the corresponding two categories; in order to ensure the accuracy of the expertise model prediction, an additional category Used to represent samples that do not belong to the two selected categories; therefore, samples related to the two categories and samples belonging to the categories are extracted from the overall data set. samples to achieve the prediction task based on ternary comparison; finally, the training dataset of this subtask is the combination of the above three datasets; S22: Expertise model training: The expertise model is responsible for performing a simple three-way comparison task to determine whether the mechanism of a drug-target pair belongs to class 1, class 2, or Class; Each expertise model is based on the classic encoder-decoder architecture; The encoder module includes two encoders, which process drug molecules and target proteins respectively, thereby obtaining representation vectors of drugs and proteins: 1) The drug molecule m is represented as a molecular graph, and then the naive GCN model is used to iteratively update its initial biochemical feature representation Specifically, the node embedding of layer l is Updated to: in is the output node embedding of the (l-1)th layer; is the trainable weight of the lth layer; A is the adjacency matrix of the molecular graph; σ(·) represents the nonlinear activation layer; after L iterative calculations, the final node embedding vector can be obtained Finally, a global pooling layer is applied to obtain the representation of the entire drug molecule graph; 2) The target protein t is represented as an amino acid sequence of unequal length, and then a 1D CNN model is used to iteratively update its initial biochemical feature representation Specifically, the node embedding of layer l is Updated to: in is the output embedding of the (l-1)th layer; k (l) Represent the number of input channels, output channels and convolution kernel size of the lth layer respectively; σ(·) represents the nonlinear activation layer; after K iterative calculations, the final protein representation can be obtained. The decoder module takes the representation vector of the combination of drugs and proteins from the encoder as input and uses multiple linear layers to obtain a 3D output, where each dimension represents 1, 2, or Class, the probability score of the node is obtained through the Softmax nonlinear layer The prediction of the category probability score is optimized using the cross entropy classification loss function. The specific formula is as follows: Where y is the true label of the sample. If the interaction between the samples is of type 1, the value is 0; if the interaction is of type 2, the value is 1; if the interaction between the samples is neither of type 1 nor of type 2, the value is -1.

4. The method for predicting drug-target interaction mechanism based on ternary comparison professional knowledge decision-making according to claim 1, characterized in that: Step S3 specifically includes: After obtaining the prediction labels of the subtasks from all the expertise models, a decision voting strategy based on class balance is adopted to effectively integrate the prediction results of each expertise model; specifically, C*(C-1) / 2 initial prediction results Q are obtained from the expertise models, which are defined as follows: Q=(q 12 ,q 13 ,…,q 1C ,q 23 ,…,q 2C ,…,q C-1,C ), (4) where q i,j represents the output of the ternary comparison task between the i-th and j-th categories, and q i,j ∈{-1,0,1}; Next, based on the three possible outputs, update the final voting vector Y∈R of C categories C As follows: 1) If q i,j is 0, indicating that category i is the output label, and the reward will be assigned to the voting score of category i, denoted as Y i ; 2) If q i,j is 1, indicating that category j is the output label, and the reward will be assigned to the voting score of category j, denoted as Y j ; 3) If q i,j -1, indicating category is the output label, and the penalty will be assigned to Y i and Y j ; Finally, after traversing the initial predictions of all professional models, the voting scores of all categories are counted; due to the imbalance between categories, a class balance weight vector H∈R based on the number of samples is used C To calculate the final score, the category with the most votes is selected as the final prediction result.

5. The method for predicting drug-target interaction mechanism based on ternary comparison professional knowledge decision-making as claimed in claim 1, characterized in that: Step S4 specifically includes: All samples in the drug-target interaction mechanism dataset are divided into training set, validation set and test set according to the ratio of 80%, 10% and 10% respectively. The data are divided into batches and sent to the model for multiple Epoch training. The adaptive learning rate optimization algorithm is used to iteratively update the model parameters. In this process, the performance of each Epoch model is recorded, and the performance of the model is evaluated by using the performance of the accuracy and F1 value. The model parameters with the best average performance are selected to finally obtain the drug-target interaction mechanism prediction model.

6. A drug-target interaction mechanism prediction system based on ternary comparison expertise decision-making, characterized by: It comprises a memory and one or more processors, wherein the memory stores executable code, and when the one or more processors execute the executable code, it is used to implement the drug-target interaction mechanism prediction method based on ternary comparison professional knowledge decision as described in any one of claims 1-5.

7. A computer-readable storage medium, characterized in that: A program is stored thereon, and when the program is executed by a processor, the drug-target interaction mechanism prediction method based on ternary comparison professional knowledge decision-making described in any one of claims 1 to 5 is implemented.

Images