A drug interaction prediction method based on DS evidence theory

Abstract

The application discloses a drug interaction prediction method based on DS evidence theory. First, multi-modal data of drugs are acquired to construct a drug interaction dataset; then, multi-modal features of the drugs are extracted as inputs of a deep neural network, a multi-modal drug interaction prediction model is constructed, and the prediction model is trained; then, the trained prediction model is used to predict drug interactions. The prediction model uses Dirichlet distribution to calculate classification probability and uncertainty value, and obtains multi-modal prediction results and overall uncertainty value through Dempster-Shafer evidence theory fusion. Compared with existing prediction methods, the application uses multi-modal drug data, has a lower prediction error, can provide meaningful uncertainty information for guiding multi-modal prediction fusion and expressing prediction confidence, and has higher reliability and robustness.

Description

Technical Field

[0001] This invention belongs to the fields of bioinformatics and computer applications, and specifically relates to a drug interaction prediction method based on DS evidence theory. Background Technology

[0002] Drug combination therapy is a common treatment approach, with many patients taking multiple medications simultaneously to treat complex or comorbid diseases. However, the main threat to patients using drug combination therapy is the increased risk of side effects due to drug-drug interactions. When one drug is taken concurrently with another, if the activity of one or both drugs changes, it means that the combination therapy can lead to an excessive response in the patient, exceeding the expected effect without predicting drug interactions. In clinical practice, the analysis and prediction of drug interaction mechanisms can encourage medical professionals to make more rational drug combinations, standardize medication practices, prevent irrational drug use, and significantly improve therapeutic efficacy, leading to faster improvement in patients' conditions and symptoms, especially for elderly patients with various chronic diseases and those requiring long-term maintenance therapy. Furthermore, the study and prediction of drug interactions can avoid or reduce toxic side effects caused by drug interactions, improving the treatment benefit / cost ratio. This facilitates the efficient development of new drugs, thereby reducing development costs, minimizing waste of medical resources, and reducing the burden on patients and society. The side effects of a given drug combination can be clinically manifested; however, given the diversity of drug combinations, it is nearly impossible to detect all possible drug pairs. Experimental screening of drug pairings presents a significant challenge in terms of cost and time. Therefore, predicting drug interactions through computational methods is of great research importance.

[0003] Existing computational methods typically focus only on the structural features of drugs, using structural similarity to predict drug-drug interactions or construct drug relationship networks, and employing graph embedding methods to predict drug-drug interactions. These methods only consider single drug molecule structural features, while drug interactions originate from multiple aspects, including drug molecule structure, drug targets, and the effects of drug-metabolizing enzymes. Therefore, existing computational methods suffer from inaccuracies and unreliability. This invention provides a drug interaction prediction method based on Dempster-Shafer (DS) evidence theory. It utilizes deep learning technology to learn drug characterization from a large amount of drug data for prediction, and then employs a decision fusion rule based on Dempster-Shafer (DS) evidence theory to obtain more reliable multimodal drug interaction prediction results. Summary of the Invention

[0004] The purpose of this invention is to improve upon the shortcomings of existing technologies that use single drug characteristics and have unreliable prediction results, and to provide a drug interaction prediction method based on DS evidence theory, which integrates multimodal drug data to predict drug interactions.

[0005] The present invention provides a drug interaction prediction method based on DS evidence theory, which mainly includes the following steps:

[0006] Step S1. Data Collection: Obtain multimodal heterogeneous drug data and drug interaction data from multiple databases such as DrugBank and UniProt to construct a drug interaction dataset;

[0007] Step S2. Data processing: Convert the multimodal data of the drug into a computable binary feature representation, and concatenate the features of the two drugs as the feature vector of the drug interaction event;

[0008] Step S3. Model Building: Construct a multimodal drug interaction prediction model, and use the drug interaction dataset constructed in step S1 as data samples and the feature vectors of the drug interaction events in step S2 as input to train the prediction model, and obtain the trained drug interaction prediction model.

[0009] Step S4. Interaction Prediction: Use the prediction model trained in Step S3 to predict drug-drug interaction events.

[0010] The multimodal heterogeneous data of the drug obtained in step S1 mainly includes the drug's molecular structure data, drug target data, and drug metabolic enzyme data.

[0011] The step S2, which converts the heterogeneous data of the drug into a computable binary feature representation, includes: using the Morgan molecular fingerprint algorithm to convert the SMILE symbol representation of the drug molecular structure into a binary feature representation; using 0 and 1 to represent the presence or absence of the drug target and drug-metabolizing enzymes, and also converting the drug target features and metabolic enzyme features into computable binary feature representations.

[0012] The multimodal drug interaction prediction model trained in step S3 mainly includes three modules: a drug multimodal feature extraction module, a deep learning network module, and a DS evidence theory decision fusion module. The drug multimodal feature extraction module includes a drug molecular structure feature extraction sub-module, a drug target feature extraction module, and a drug metabolic enzyme feature extraction sub-module. The deep learning network module includes a drug molecular structure prediction sub-network module, a drug target prediction sub-network prediction sub-network module, and a drug metabolic enzyme prediction sub-network module. The drug multimodal feature extraction module is connected to the deep learning network module, and the deep learning network module is connected to the DS evidence theory decision fusion module.

[0013] The drug multimodal feature extraction module includes a drug molecular structure feature extraction submodule, a drug target feature extraction submodule, and a drug metabolic enzyme feature extraction submodule. Principal component analysis (PCA) algorithm is used to reduce the dimensionality of the feature vectors of different modalities and extract low-dimensional features as input to the deep learning network module.

[0014] The deep learning network module is used to generate prediction results of drug interactions. It includes a drug molecule structure prediction sub-network module, a drug target prediction sub-network module, and a drug metabolic enzyme prediction sub-network module. Each sub-network module uses different modal feature vectors of the drug as input, and then calculates the prediction results.

[0015] The DS evidence theory decision fusion module uses the prediction results calculated by the deep learning network module as parameters of the Dirichlet distribution to calculate the uncertainty of a single mode. Then, based on the uncertainty of each mode, the DS evidence theory is used to fuse the prediction results of the multimodal modes to obtain the final prediction result and the overall uncertainty.

[0016] Compared with the prior art, the present invention has the following beneficial effects and advantages:

[0017] (1) Obtain multimodal heterogeneous data of drugs and effectively improve the reliability of prediction by utilizing the correlation and complementarity between different drug modalities.

[0018] (2) The introduction of the Dirichlet distribution makes the uncertainty of prediction for each modality accurately estimated, thus improving the robustness of classification.

[0019] (3) Using a fusion strategy based on DS evidence theory to fuse the prediction results of drug multimodal features can improve the accuracy of classification and clearly express the confidence level of the prediction. Attached Figure Description

[0020] Figure 1 This diagram illustrates the basic steps of a drug interaction prediction method based on DS evidence theory according to the present invention.

[0021] Figure 2 This is a schematic diagram of the structure of the drug interaction prediction model trained in step S3 of this invention.

[0022] The diagram is labeled as follows: 1. Drug multimodal feature extraction module; 1-1. Drug molecular structure feature extraction sub-module; 1-2. Drug target feature extraction module; 1-3. Drug metabolic enzyme feature extraction sub-module; 2. Deep learning network module; 2-1. Drug molecular structure prediction sub-network module; 2-2. Drug target prediction sub-network prediction module; 2-3. Drug metabolic enzyme prediction sub-network module; 3. DS evidence theory decision fusion module. Detailed Implementation

[0023] Example:

[0024] like Figure 1 As shown, the embodiments of the present invention include the following four steps:

[0025] Step S1. Data collection: Obtain drug multimodal heterogeneous data and drug interaction data from multiple databases such as DrugBank and UniProt, construct a drug interaction dataset, and obtain drug multimodal heterogeneous data, which mainly includes drug molecular structure data, drug target data, and drug metabolic enzyme data;

[0026] Step S2. Data processing: Converting the multimodal data of the drug into a computable binary feature representation, including: using the Morgan molecular fingerprint algorithm to convert the SMILE symbol representation of the drug molecular structure into a binary feature representation; using 0 and 1 to represent the presence or absence of the drug target and drug-metabolizing enzyme, and converting the drug target features and metabolic enzyme features into computable binary feature representations; and then concatenating the features of the two drugs as a feature vector of the drug interaction event.

[0027] Step S3. Model Building: Construct a multimodal drug interaction prediction model, and use the drug interaction dataset constructed in step S1 as data samples and the feature vectors of the drug interaction events in step S2 as input to train the prediction model, and obtain the trained drug interaction prediction model.

[0028] Step S4. Effect prediction: Use the prediction model obtained in step S3 to predict drug interaction events.

[0029] like Figure 2 As shown, the drug interaction prediction model trained in step S3 of this embodiment mainly includes three modules: a drug multimodal feature extraction module 1, a deep learning network module 2, and a DS evidence theory decision fusion module 3. The drug multimodal feature extraction module 1 includes a drug molecule structure feature extraction submodule 1-1, a drug target feature extraction module 1-2, and a drug metabolic enzyme feature extraction submodule 1-3; the deep learning network module 2 includes a drug molecule structure prediction subnetwork module 2-1, a drug target prediction subnetwork prediction module 2-2, and a drug metabolic enzyme prediction subnetwork module 2-3.

[0030] The drug multimodal feature extraction module 1 includes drug molecular structure feature extraction submodules 1-2, drug target feature extraction submodules 1-2, and drug metabolic enzyme feature extraction submodules 1-3. Principal component analysis (PCA) algorithm is used to reduce the dimensionality of feature vectors of different modalities and extract low-dimensional features as input to deep learning network modules.

[0031] The deep learning network module 2 is used to generate prediction results for drug interactions. It includes a drug molecule structure prediction sub-network module 2-1, a drug target prediction sub-network module 2-2, and a drug metabolic enzyme prediction sub-network module 2-3, each using different modal features of the drug as input, and then calculating the prediction results. A preferred formula for the prediction sub-network is shown below:

[0032] f(x)=f[a (3) (h (2) (a (2) (h (1) (a (1) (x)))))],a (l) (x)=W (l) c+b (l) h (l) (c) = max(0,x) (1)

[0033] In the formula, x represents the eigenvector of the drug interaction event, and W... (l) and b (l) This represents the weights and biases of the l-th hidden layer that need to be calculated, where l takes the values ​​1 and 2.

[0034] The DS evidence theory decision fusion module 3 uses the result vector 'a' predicted by the sub-network module in the deep learning network module 2 as the parameter of the Dirichlet distribution to calculate the probability p corresponding to different interaction events. k For a class K drug interaction event, and an uncertainty value u, the specific calculation formula is as follows:

[0035]

[0036] In the formula, a k This represents the k-th value of vector a. Let represent the Dirichlet strength, and u represent the prediction uncertainty. Then, the predictions from multiple sub-network modules are fused using Dirichlet evidence theory to obtain the final prediction result and the overall uncertainty value. The specific calculation formula is as follows:

[0037]

[0038] in the formula Used to measure the degree of conflict between category probability predictions.

[0039] In summary, this invention relates to a drug interaction prediction method based on Dirichlet distribution (DS) evidence theory. This method fully utilizes the correlation and complementarity between different modal characteristics of drugs to predict potential drug interactions. It quantifies the uncertainty of features using Dirichlet distribution and DS evidence theory, and then uses this uncertainty information to guide the fusion of multimodal predictions. Compared with existing prediction methods, this invention not only has lower prediction errors but also provides meaningful uncertainty information to guide multimodal prediction fusion and express prediction confidence, resulting in higher accuracy and reliability.

[0040] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A drug interaction prediction method based on DS evidence theory, characterized in that, Includes the following steps: Step S1. Data Collection: Obtain drug multimodal heterogeneity data and drug interaction data from DrugBank and UniProt databases to construct a drug interaction dataset; Step S2. Data processing: Convert the multimodal data of the drug into a computable binary feature representation, and concatenate the features of the two drugs as the feature vector of the drug interaction event; Step S3. Model Building: Construct a multimodal drug interaction prediction model, and use the drug interaction dataset constructed in step S1 as data samples and the feature vectors of the drug interaction events in step S2 as input to train the prediction model, and obtain the trained drug interaction prediction model. Step S4. Interaction Prediction: Use the prediction model trained in Step S3 to predict drug-drug interaction events; The multimodal drug interaction prediction model in step S3 mainly includes three modules: a drug multimodal feature extraction module, a deep learning network module, and a DS evidence theory decision fusion module; the drug multimodal feature extraction module includes a drug molecular structure feature extraction submodule, a drug target feature extraction module, and a drug metabolic enzyme feature extraction submodule; The deep learning network module includes a drug molecule structure prediction sub-network module, a drug target prediction sub-network module, and a drug metabolic enzyme prediction sub-network module; the drug multimodal feature extraction module is connected to the deep learning network module, and the deep learning network module is connected to the DS evidence theory decision fusion module. The deep learning network module is used to generate prediction results of drug interactions. The drug molecule structure prediction sub-network module, the drug target prediction sub-network module, and the drug metabolic enzyme prediction sub-network module use different modal feature vectors of the drug as inputs, and then calculate the prediction results. The DS evidence theory decision fusion module uses the prediction results calculated by the deep learning network module as parameters of the Dirichlet distribution to calculate the uncertainty of a single mode. Then, based on the uncertainty of each mode, the DS evidence theory is used to fuse the prediction results of the multimodal modes to obtain the final prediction result and the overall uncertainty.

2. The drug interaction prediction method based on DS evidence theory according to claim 1, characterized in that, The multimodal heterogeneous data of the drug obtained in step S1 mainly includes the drug's molecular structure data, drug target data, and drug metabolic enzyme data.

3. The drug interaction prediction method based on DS evidence theory according to claim 1, characterized in that, The step S2, which converts the heterogeneous data of the drug into a computable binary feature representation, includes: using the Morgan molecular fingerprint algorithm to convert the SMILE symbol representation of the drug molecular structure into a binary feature representation; using 0 and 1 to represent the presence or absence of the drug target and drug-metabolizing enzymes, and also converting the drug target features and metabolic enzyme features into computable binary feature representations.

4. The drug interaction prediction method based on DS evidence theory according to claim 1, characterized in that, The drug multimodal feature extraction module includes a drug molecular structure feature extraction submodule, a drug target feature extraction submodule, and a drug metabolic enzyme feature extraction submodule. Principal component analysis (PCA) algorithm is used to reduce the dimensionality of feature vectors of different modalities and extract low-dimensional features as input to the deep learning network module.

Images