Small sample drug interaction prediction method fusing drug association relationship
By using graph attention network and Transformer encoder to extract the characteristics of drug molecules and combining the model-independent meta-learning framework, the accuracy problem of drug-drug interaction prediction in small sample scenarios is solved, and more efficient drug interaction prediction is achieved.
Patent Information
- Application Number
- CN202411815872.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-11
- Publication Date
- 2025-04-25
AI Technical Summary
The prior art is difficult to accurately predict drug-drug interactions in small sample scenarios, especially in issues such as scarcity of data, disease heterogeneity and individual differences.
The graph attention network GAT is used to extract the features of drug molecules, and the features are aggregated through the Transformer encoder, combined with the model-independent learning MAML framework, and fuse the DDI similarity relationship for small sample learning.
It improves the accuracy of drug-drug interaction prediction, optimizes the update direction of model parameters, and is suitable for small sample scenarios.
Smart Images
Figure CN119889523A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of drug interaction prediction, and in particular to a small sample drug interaction prediction method integrating drug association relationships. Background Art
[0002] Drug-drug interaction (DDI) prediction refers to the process of predicting and evaluating possible interactions between different drugs. Its purpose is to identify the pharmacokinetic and pharmacodynamic changes that may occur when multiple drugs act together. Nowadays, it is very common to use multiple drugs to treat different diseases or complex diseases. How to match drugs is related to the safety and efficacy of patients. Therefore, by predicting drug-drug interactions, it can help medical staff better carry out drug treatment, while improving the treatment effect of patients and reducing the risk of adverse reactions caused by combined medication. At the same time, with the emergence of computational methods and the explosive growth of biological data, researchers have begun to use computer models and data-driven technologies to process and analyze large amounts of biological data to conduct efficient and low-cost drug interaction studies.
[0003] How to represent drug molecules and then extract feature representations that can express drug properties is the primary task of this study. Since molecules can be represented as graphs with atoms as nodes and chemical bonds as edges, using graph neural networks to extract molecular features and then training prediction models through deep learning algorithms has become a hot topic in current research. The SSI-DDI model takes into account the important role of molecular substructures in DDI prediction, directly learns the feature representations of these substructures from the molecular graph, and then identifies pairwise interactions between two drug substructures. The SA-DDI model proposes a substructure-aware graph neural network designed to capture drug substructures with adaptive size and shape. The DGNN-DDI model designs a directed message passing neural network with a substructure attention mechanism that can identify interacting substructures, thereby enhancing the final characteristics of the drug. In order to overcome the limitation of GNN in capturing long-distance dependencies, some transformer-based methods have also been tried in drug feature extraction. The MolGNet model uses a transformer with a GRU network to aggregate information and update the state of nodes in the molecular graph. The SSR-DDI model designs a substructure enhancement module using a graph transformer to capture the structural information of drug molecules.
[0004] However, in actual research, we often face problems such as data scarcity, disease heterogeneity, and individual differences, which are fatal to traditional deep learning models that rely on a large number of training samples. Studies have shown that the data distribution of different DDIs is extremely unbalanced, and there is only a small amount of labeled data for a large number of DDI types.
[0005] Therefore, there is a need for a small sample drug interaction prediction method that integrates drug association relationships and can accurately predict drug-drug interactions in small sample scenarios. Summary of the invention
[0006] The main purpose of the present invention is to provide a small sample drug interaction prediction method integrating drug association relationships to solve the problem that the prior art cannot accurately predict drug-drug interactions in small sample scenarios.
[0007] To achieve the above object, the present invention provides a small sample drug interaction prediction method integrating drug association relationships, which specifically comprises the following steps:
[0008] S1, uses the graph attention network GAT to extract the features of two drug molecules, aggregates the features of the two drug molecules, and extracts the features of drug-drug interaction.
[0009] S2, the extracted drug-drug interaction features are used as DDI cache features, and the DDI similarity relationship is calculated by calculating the feature distance.
[0010] S3, based on the model-independent meta-learning MAML framework, integrates the DDI similarity relationship for small sample learning.
[0011] Furthermore, step S1 specifically includes the following steps:
[0012] S1.1, for a pair of drug molecules (g1, g2), GAT extracts the feature vectors of each layer of the two drug molecules through L message passing layers and is the L-th layer feature vector of drug molecule g1, is the L-th layer feature vector of drug molecule g2.
[0013] S1.2, concatenate the feature vectors of each layer of the drug molecule pair to obtain a concatenated vector
[0014] S1.3, the concatenated vector is input into the Transformer encoder, and finally the output features of the Transformer encoder are aggregated as the final features of the drug pair interaction.
[0015] Furthermore, step S2 specifically includes the following steps:
[0016] S2.1, save K positive samples from N different DDI types as DDI cache features in, Indicates the Nth DDI type eN There are K positive samples in .
[0017] S2.2, using DDI class prototype Uniquely represents a DDI cache feature. It means taking the average of K positive samples.
[0018] S2.3, divide the DDI cache features into training and test classes, and calculate the similarity score s(e) between the DDI prototype of the training class and the DDI prototype of the test class i ,e t ):
[0019]
[0020] in, To test the DDI class prototype features, is the prototype feature of the training DDI class, and exp is the exponential function.
[0021] Furthermore, step S3 specifically includes the following steps:
[0022] S3.1, in the meta-training phase, the meta-learner learns a good initialization parameter θ from multiple meta-training tasks, for each task T τ , the base learner uses the loss on the support set Through gradient descent To update the initialization parameter θ, we get the meta-learner parameter θ′ τ :
[0023]
[0024] Among them, α is a learnable parameter, which represents the step size of gradient update.
[0025] S3.2, the base learner in the meta-training phase uses the meta-learner parameters θ′ τ , calculate the loss of the query set By integrating N t training tasks to update the meta-learner parameters and obtain the base learner parameters θ in the meta-training phase ＊ :
[0026]
[0027] Among them, τ is the task, argmin θ It represents the parameter θ that minimizes the objective loss function.
[0028] S3.3, using the base learner parameters θ in the meta-training phase ＊, as the initialization parameters of the meta-learner in the meta-test phase, and use the DDI sample enhancement module to calculate the loss of the query set, and use the loss of the query set to update the base learner parameters in the meta-test phase.
[0029] S3.4, using the base learner parameters in the meta-testing phase to predict small sample drug interactions.
[0030] Furthermore, the following calculations are performed in the DDI sample enhancement module:
[0031] After step S2.3, the DDI cache feature similar to the DDI cache feature of the test class is calculated, that is, the similar DDI cache feature is averaged to obtain the intermediate feature h m ;
[0032] Update the DDI cache characteristics of the support set on the test class through formula (4):
[0033]
[0034] in, To enhance the features of the test class DDI, Test class DDI cache features.
[0035] The present invention has the following beneficial effects:
[0036] The present invention incorporates a small sample learning method of DDI association, uses similar DDI cache features to enhance the sample features of the test-type DDI, and thus optimizes the update direction of the model parameters. The use of graph neural networks can more efficiently mine drug molecule information and improve the accuracy of drug-drug interaction prediction. BRIEF DESCRIPTION OF THE DRAWINGS
[0037] In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the following will briefly introduce the drawings required for the specific implementation or the prior art description. Obviously, the drawings described below are some implementations of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work. In the drawings:
[0038] Figure 1 A flow chart showing step S1 of the present invention is shown.
[0039] Figure 2 A flow chart showing step S3 of the present invention is shown. DETAILED DESCRIPTION
[0040] The technical solution of the present invention will be described clearly and completely below in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0041] A small sample drug interaction prediction method integrating drug association relationships specifically comprises the following steps:
[0042] S1, uses the graph attention network GAT to extract the features of two drug molecules, aggregates the features of the two drug molecules, and extracts the features of drug-drug interaction.
[0043] S2, the extracted drug-drug interaction features are used as DDI cache features, and the DDI similarity relationship is calculated by calculating the feature distance.
[0044] S3, based on the model-independent meta-learning MAML framework, integrates the DDI similarity relationship for small sample learning.
[0045] Specifically, Figure 1 As shown in Figure 1, the interaction between drugs is generally determined by some special structures in the drug molecules. Considering that general graph neural networks do not pay special attention to these substructures, the present invention uses the graph attention network GAT to extract the features of drug molecules. It regards different neighbor nodes as having different importance and can use the attention mechanism to learn these special structural information during the message passing process. For a pair of drug molecules (g1, g2), GAT extracts the feature vectors of each layer through L message passing layers. and Afterwards, in order to effectively extract the interaction and relationship between the two molecules, the feature information of the molecules is fused globally while ensuring that the local information of the molecules is not lost. The feature vectors of each layer of the spliced molecule pair of the present invention are: And input it into the Transformer encoder, and finally aggregate the output features of the Transformer as the final features of the drug pair interaction.
[0046] Step S1 specifically includes the following steps:
[0047] S1.1, for a pair of drug molecules (g1, g2), GAT extracts the feature vectors of each layer of the two drug molecules through L message passing layers and is the L-th layer feature vector of drug molecule g1, is the L-th layer feature vector of drug molecule g2.
[0048] S1.2, concatenate the feature vectors of each layer of the drug molecule pair to obtain a concatenated vector
[0049] S1.3, the concatenated vector is input into the Transformer encoder, and finally the output features of the Transformer encoder are aggregated as the final features of the drug pair interaction.
[0050] Specifically, step S2 includes the following steps:
[0051] S2.1, Generally speaking, similar feature representations are adjacent in the feature space, and their similarity can be calculated by calculating the distance between the feature representations. In each round of training, K positive samples from N different DDI types are saved as DDI cache features. in, Indicates the Nth DDI type e N There are K positive samples in .
[0052] S2.2, using DDI class prototype Uniquely represents a DDI cache feature. It means taking the average of K positive samples.
[0053] S2.3, the DDI cache features are divided into training and test classes, in order to calculate the DDI prototype of the training class DDI class prototype with test class The present invention uses formula (1) to convert the difference in distance between feature representations into a similarity score s(e i ,e t ):
[0054]
[0055] in, To test the DDI class prototype features, is the prototype feature of the training DDI class, and exp is the exponential function.
[0056] Specifically, meta-learning is one of the most popular small sample learning methods. The present invention adopts a meta-learning framework based on MAML, and then integrates the association relationship of DDI into MAML to improve the prediction accuracy of the prediction model.
[0057] like Figure 2 As shown, step S3 specifically includes the following steps:
[0058] S3.1, in the meta-training phase, the meta-learner learns a good initialization parameter θ from multiple meta-training tasks, for each task T τ, the base learner uses the loss on the support set Through gradient descent To update the initialization parameter θ, we get the meta-learner parameter θ′ τ :
[0059]
[0060] Among them, α is a learnable parameter, which represents the step size of gradient update.
[0061] S3.2, the base learner in the meta-training phase uses the meta-learner parameters θ′ τ , calculate the loss of the query set By integrating N t training tasks to update the meta-learner parameters and obtain the base learner parameters θ in the meta-training phase ＊ :
[0062]
[0063] Among them, τ is the task, argmin θ It represents the parameter θ that minimizes the objective loss function.
[0064] S3.3, using the base learner parameters θ in the meta-training phase ＊ , as the initialization parameters of the meta-learner in the meta-test phase, and use the DDI sample enhancement module to calculate the loss of the query set, and use the loss of the query set to update the base learner parameters in the meta-test phase.
[0065] S3.4, using the base learner parameters in the meta-testing phase to predict small sample drug interactions.
[0066] During the meta-testing process, the meta-learner has learned a good initialization parameter θ ＊ , and further fine-tuning will be performed on the support set of the meta-test phase to adapt to the new task. The support set contains K drug pairs that can interact with each other and K drug pairs that cannot interact with each other. After fine-tuning, MAML can predict whether the drug pairs in the query set can interact with each other. Due to the randomness of sampling, the drug pair features sampled from the test class cannot accurately represent the actual distribution of this type of DDI features. If the sampled drug pair features deviate far from the center of the actual feature distribution, it will affect the update direction of the model parameters. Therefore, the present invention attempts to integrate the association relationship of DDI into the model during the fine-tuning process, and designs a DDI sample feature enhancement module to enhance the DDI features on the test class support set with similar type DDI cache features. After calculating the similarity of DDI using the DDI similarity relationship calculation method based on feature distance, all similar DDI cache features are averaged to obtain an intermediate feature h m, and then update the DDI features of the support set in the test class using formula (4).
[0067] Specifically, the following calculations are performed in the DDI sample enhancement module:
[0068] After step S2.3, the DDI cache feature similar to the DDI cache feature of the test class is calculated, that is, the similar DDI cache feature is averaged to obtain the intermediate feature h m ;
[0069] Update the DDI cache characteristics of the support set on the test class through formula (4):
[0070]
[0071] in, To enhance the features of the test class DDI, Test class DDI cache features.
[0072] Of course, the above description is not a limitation of the present invention, and the present invention is not limited to the above examples. Changes, modifications, additions or substitutions made by technicians in this technical field within the essential scope of the present invention should also fall within the protection scope of the present invention.
Claims
1. A small sample drug interaction prediction method integrating drug association relationships, characterized in that: The specific steps include: S1, using the graph attention network GAT to extract the features of two drug molecules, aggregate the features of the two drug molecules, and extract the features of drug-drug interaction; S2, taking the extracted drug-drug interaction features as DDI cache features, and calculating the DDI similarity relationship by calculating the feature distance; S3, based on the model-independent meta-learning MAML framework, integrates the DDI similarity relationship for small sample learning.
2. A small sample drug interaction prediction method integrating drug association relationships according to claim 1, characterized in that: Step S1 specifically includes the following steps: S1.1, for a pair of drug molecules (g1, g2), GAT extracts the feature vectors of each layer of the two drug molecules through L message passing layers and is the L-th layer feature vector of drug molecule g1, is the L-th layer feature vector of drug molecule g2; S1.2, concatenate the feature vectors of each layer of the drug molecule pair to obtain a concatenated vector S1.3, the concatenated vector is input into the Transformer encoder, and finally the output features of the Transformer encoder are aggregated as the final features of the drug pair interaction.
3. A small sample drug interaction prediction method integrating drug association relationships according to claim 2, characterized in that: Step S2 specifically includes the following steps: S2.1, save K positive samples from N different DDI types as DDI cache features in, Indicates the Nth DDI type e N There are K positive samples in ; S2.2, using DDI class prototype Uniquely represents a DDI cache feature. Indicates taking the average of K positive samples; S2.3, divide the DDI cache features into training and test classes, and calculate the similarity score s(e) between the DDI prototype of the training class and the DDI prototype of the test class i ,e t ): in, To test the DDI class prototype features, is the prototype feature of the training DDI class, and exp is the exponential function.
4. A small sample drug interaction prediction method integrating drug association relationships according to claim 3, characterized in that: Step S3 specifically includes the following steps: S3.1, in the meta-training phase, the meta-learner learns a good initialization parameter θ from multiple meta-training tasks, for each task T τ , the base learner uses the loss on the support set Through gradient descent To update the initialization parameter θ, we get the meta-learner parameter θ′ τ : Among them, α is a learnable parameter, which represents the step size of gradient update; S3.2, the base learner in the meta-training phase uses the meta-learner parameters θ′ τ , calculate the loss of the query set By integrating N t training tasks to update the meta-learner parameters and obtain the base learner parameters θ in the meta-training phase ＊ : Among them, τ is the task, argmin θ Indicates the parameter θ that minimizes the objective loss function; S3.3, using the base learner parameters θ in the meta-training phase ＊ , as the initialization parameters of the meta-learner in the meta-test phase, and use the DDI sample enhancement module to calculate the loss of the query set, and use the loss of the query set to update the base learner parameters in the meta-test phase; S3.4, using the base learner parameters in the meta-testing phase to predict small sample drug interactions.
5. A small sample drug interaction prediction method integrating drug association relationships according to claim 4, characterized in that: The following calculations are performed in the DDI sample enhancement module: After step S2.3, the DDI cache feature similar to the DDI cache feature of the test class is calculated, that is, the similar DDI cache feature is averaged to obtain the intermediate feature h m ; Update the DDI cache characteristics of the support set on the test class through formula (4): in, To enhance the features of the test class DDI, Test class DDI cache features.