A drug interaction prediction method based on multidimensional features
By adopting the method of feature importance assessment and model integration in drug interaction prediction, the problem of excessive false positive results in the prior art is solved, the accuracy and reliability of the prediction are improved, and unnecessary treatment plan adjustments and medical costs are reduced.
Patent Information
- Application Number
- CN202411754956.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-12-03
- Publication Date
- 2025-03-21
- Estimated Expiration
- 2044-12-03
AI Technical Summary
Existing drug interaction prediction methods based on multidimensional characteristics may produce more false positive results, resulting in excessive caution of physicians, may unnecessarily adjust treatment plans, increase patient medical costs, and may delay the optimal treatment timing.
By collecting multi-dimensional feature information of drugs from authoritative databases and combining real clinical case data, the feature importance evaluation algorithm in machine learning is used to calculate the relative importance of each drug multi-dimensional feature, assign corresponding weights, construct multiple drug interaction prediction models, and integrate the prediction results of multiple models, and obtain the final prediction results using the weighted average method.
It reduces false positive results due to unreasonable feature weights, enhances the generalization ability of the model, improves the applicability and reliability of drug interaction predictions, and reduces the probability of a single model producing false positive results.
Smart Images
Figure CN119230128B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of drug interaction prediction, in particular to a drug interaction prediction method based on multidimensional features. Background Art
[0002] In actual medical scenarios, doctors will refer to drug interaction prediction results when formulating treatment plans for patients; however, drug interaction prediction methods based on multidimensional features may produce more false positive results due to various reasons; for example, a patient with hypertension and diabetes needs to take antihypertensive and hypoglycemic drugs at the same time; when the doctor uses this prediction method, he finds that the prediction results show that the two drugs interact with each other, but after further consulting detailed information and combining clinical experience, he finds that such interactions rarely occur in practice or have little impact; too many false positive results will cause doctors to be overly cautious and may adjust treatment plans unnecessarily, increase patients' medical costs, such as changing drugs, adding additional monitoring, etc., and may even delay the best treatment time for patients because doctors spend time on troubleshooting false positive results instead of giving patients the most suitable treatment in time. Summary of the invention
[0003] In order to solve the above technical problems, the present invention is implemented by the following technical solutions: a drug interaction prediction method based on multidimensional features, comprising the following steps:
[0004] S1: Collect multidimensional feature information of drugs from authoritative databases such as DrugBank, KEGG, and Pubchem to obtain drug features; collect real clinical case data, record them as clinical data, analyze the features that are closely related to the actual interaction between clinical data and drug features, remove drug features with weak correlation or prone to misjudgment, screen out features that are closely related to drug interactions, and obtain multidimensional drug features;
[0005] S2: Based on the multidimensional features of drugs obtained after screening, the feature importance evaluation algorithm in machine learning is used to calculate the relative importance of each multidimensional feature of drugs in predicting drug interactions and obtain the importance index; according to the importance index of the multidimensional features of drugs, a corresponding weight is assigned to each feature to obtain the weight assignment result; so that important features can play a greater role in the prediction model and reduce false positive results caused by unreasonable feature weights;
[0006] S3: Based on the multidimensional characteristics of drugs, multiple drug interaction prediction models are constructed, and the prediction results of multiple prediction models are integrated. The weighted average method is used to assign different weights to each prediction result according to the accuracy and reliability of each model to obtain the final prediction result;
[0007] S4: Compare the final prediction result with the collected real-world drug use data to obtain a comparison result, and provide the final prediction result to the clinician based on the comparison result.
[0008] Preferably, the multidimensional characteristics of the drug include the chemical structure characteristics, pharmacological action characteristics, metabolic pathway characteristics and gene expression characteristics of the drug.
[0009] Preferably, step S1 includes cooperating with medical institutions to collect real-world drug use data, which is recorded as real data. The real data includes actual reactions of patients after taking multiple drugs at the same time and adverse reaction monitoring data.
[0010] Preferably, in step S2, the relative importance of each drug multidimensional feature in predicting drug interactions is calculated to obtain an importance index, and the process is as follows:
[0011] Based on the multidimensional characteristics of the drugs obtained after screening, the drug interaction result data corresponding to the drug samples are prepared as labels to obtain binary labels or probability values. The binary labels include interaction and no interaction. The probability value indicates the probability of interaction when the drugs interact.
[0012] Based on random forest as the feature importance evaluation algorithm, the contribution of each feature to reducing impurity when building a decision tree is calculated to evaluate the feature importance. In the case of multiple trees, the importance of the features in each tree is averaged to obtain the final feature importance evaluation result.
[0013] The random forest model is trained using drug multidimensional features and corresponding drug interaction label data to construct a random forest model, output the importance score of each feature, and sort the calculated importance results of each feature from high to low according to the importance score to obtain the importance sequence of each drug multidimensional feature in predicting drug interactions. Based on the importance sequence and combined with the preset importance index comparison list, the position of the drug feature in the importance sequence is matched with the importance index matched in the importance index comparison list, and the relative importance of different drug multidimensional features in predicting drug interactions is reflected through the importance index.
[0014] Preferably, step S2 includes a weight allocation process:
[0015] The random forest algorithm was used to construct a feature importance assessment model, with the retained multidimensional features of drugs as input variables and the occurrence or non-occurrence of drug interactions as output variables to train the random forest model;
[0016] During the training process, the information gain and Gini index reduction of each feature in the decision tree construction process were calculated, and the importance score of each feature for predicting drug interactions was determined based on the calculated information gain and Gini index reduction;
[0017] The weights are assigned according to the importance scores. The importance scores of each feature are converted into weight values by using the normalization method so that the sum of all weights is 1.
[0018] Preferably, in step S3, the multiple drug interaction prediction models include a chemical structure characteristic model, a pharmacological action characteristic model, a metabolic pathway characteristic model and a gene expression characteristic model.
[0019] Preferably, step S3 includes constructing multiple drug interaction prediction models, the process is as follows:
[0020] For the chemical structure feature model, the chemical structure features of the drug are obtained, the chemical structure network is constructed, and the influence of the chemical structure features on drug interactions is analyzed through molecular dynamics simulation to obtain the structure prediction results;
[0021] For the pharmacological action characteristic model, the pharmacological action of the drug is obtained, the drug target network is constructed, and the network pharmacology is used to analyze how the drug exerts its effects by affecting multiple targets and biological pathways, as well as the potential interactions with the drug, to obtain the pharmacological prediction results;
[0022] For drug metabolism pathway models, we collect enzyme-catalyzed reactions and metabolites during drug metabolism in the body, build a metabolic pathway network, analyze the interaction relationship between drugs in the metabolic pathway, use pathway analysis algorithms to calculate the possibility of drug interactions, and obtain metabolic prediction results;
[0023] For the gene expression characteristic model, the change data of gene expression under the action of drugs is obtained, the gene regulatory network is constructed, and the network analysis method is used to predict the impact of drug interactions on gene expression and the drug interactions caused by them, and the gene prediction results are obtained.
[0024] Preferably, step S3 further includes: based on the chemical structure characteristic model, the pharmacological action characteristic model, the metabolic pathway characteristic model and the gene expression characteristic model, for each model, dividing the collected data set according to an appropriate ratio, and evaluating the trained model using a test set based on evaluation indicators, wherein the evaluation indicators include accuracy, recall rate, F1 value and mean square error; measuring the accuracy and reliability of each model in predicting drug interactions by the evaluation indicators, and performing weight allocation based on the comparison results of the evaluation indicators of each model on the test set;
[0025] For new drug data to be predicted, they are input into the chemical structure characteristic model, pharmacological action characteristic model, metabolic pathway characteristic model and gene expression characteristic model respectively to obtain the prediction results of each model.
[0026] Preferably, for obtaining the prediction results of each model, the prediction results include binary results or probability values; the binary results indicate whether there is an interaction, and the probability values indicate the probability of the interaction occurring when the drugs have an interaction;
[0027] If the predictions were binary, the predictions of each model were converted to numerical values: 1 for interaction and 0 for no interaction;
[0028] Then, a weighted average is performed according to the determined weights, and the judgment threshold is set to 0.5. If the weighted average value is greater than the judgment threshold, the final prediction result is that there is interaction, otherwise there is no interaction;
[0029] If the prediction result is a probability value, simply multiply the prediction probability of each model by the corresponding weight and add them together to obtain the final prediction probability value, which is used as the final prediction result.
[0030] Preferably, step S1 includes:
[0031] For the multidimensional characteristics of drugs, the Pearson correlation coefficient of the occurrence and severity of the interaction between clinical data and drug characteristics is calculated to obtain the significance level and Pearson correlation coefficient; according to the size and significance level of the Pearson correlation coefficient, the characteristics with strong correlation are screened to obtain the multidimensional characteristics of drugs;
[0032] The significance level is set to A. If the p-value corresponding to the Pearson correlation coefficient of the drug feature is less than A and the absolute value of the Pearson correlation coefficient is greater than B, the feature is considered to be closely related to the drug interaction and is retained; otherwise, if the p-value is greater than A or the absolute value of the Pearson correlation coefficient is less than B, the feature is removed.
[0033] The present invention provides a drug interaction prediction method based on multidimensional features, which has the following beneficial effects:
[0034] This multi-dimensional feature-based drug interaction prediction method, through integrated learning and model fusion, takes advantage of the complementary advantages of different models. Reasonable screening and weighted features can enable the model to maintain relatively stable performance in different data sets and application scenarios, reduce model prediction bias caused by data fluctuations or feature changes, enhance the generalization ability of the model, improve its applicability and reliability in actual drug interaction prediction, and reduce the probability of a single model producing false positive results. BRIEF DESCRIPTION OF THE DRAWINGS
[0035] Figure 1The present invention is a flowchart of a drug interaction prediction method based on multidimensional features. DETAILED DESCRIPTION
[0036] The present invention is further described in detail below in conjunction with the accompanying drawings and specific embodiments. The embodiments of the present invention are provided for the purpose of illustration and description, and are not intended to be exhaustive or to limit the present invention to the disclosed forms. Many modifications and variations will be apparent to those of ordinary skill in the art. The embodiments are selected and described in order to better illustrate the principles and practical applications of the present invention, and to enable those of ordinary skill in the art to understand the present invention and thereby design various embodiments with various modifications suitable for specific uses.
[0037] like Figure 1 As shown, the present invention provides a technical solution: a drug interaction prediction method based on multidimensional features, comprising the following steps:
[0038] S1: Collect multidimensional feature information of drugs from authoritative databases such as DrugBank, KEGG, and Pubchem to obtain drug features; collect real clinical case data, record them as clinical data, analyze the features that are closely related to the actual interaction between clinical data and drug features, remove drug features with weak correlation or prone to misjudgment, screen out features that are closely related to drug interactions, and obtain multidimensional drug features;
[0039] S2: Based on the multidimensional features of drugs obtained after screening, the feature importance evaluation algorithm in machine learning is used to calculate the relative importance of each multidimensional feature of drugs in predicting drug interactions and obtain the importance index; according to the importance index of the multidimensional features of drugs, a corresponding weight is assigned to each feature to obtain the weight assignment result; so that important features can play a greater role in the prediction model and reduce false positive results caused by unreasonable feature weights;
[0040] S3: Based on the multidimensional characteristics of drugs, multiple drug interaction prediction models are constructed, and the prediction results of multiple prediction models are integrated. The weighted average method is used to assign different weights to each prediction result according to the accuracy and reliability of each model to obtain the final prediction result;
[0041] S4: Compare the final prediction result with the collected real-world drug use data to obtain a comparison result, and provide the final prediction result to the clinician based on the comparison result.
[0042] Preferably, the multidimensional characteristics of the drug include the chemical structure characteristics, pharmacological action characteristics, metabolic pathway characteristics and gene expression characteristics of the drug.
[0043] Preferably, step S1 includes cooperating with medical institutions to collect real-world drug use data, which is recorded as real data. The real data includes actual reactions of patients after taking multiple drugs at the same time and adverse reaction monitoring data.
[0044] Preferably, in step S2, the relative importance of each drug multidimensional feature in predicting drug interactions is calculated to obtain an importance index, and the process is as follows:
[0045] Based on the multidimensional characteristics of the drugs obtained after screening, the drug interaction result data corresponding to the drug samples are prepared as labels to obtain binary labels or probability values. The binary labels include interaction and no interaction. The probability value indicates the probability of interaction when the drugs interact.
[0046] Based on random forest as the feature importance evaluation algorithm, the contribution of each feature to reducing impurity when building a decision tree is calculated to evaluate the feature importance. In the case of multiple trees, the importance of the features in each tree is averaged to obtain the final feature importance evaluation result.
[0047] The random forest model is trained using drug multidimensional features and corresponding drug interaction label data to construct a random forest model, output the importance score of each feature, and sort the calculated importance results of each feature from high to low according to the importance score to obtain the importance sequence of each drug multidimensional feature in predicting drug interactions. Based on the importance sequence and combined with the preset importance index comparison list, the position of the drug feature in the importance sequence is matched with the importance index matched in the importance index comparison list, and the relative importance of different drug multidimensional features in predicting drug interactions is reflected through the importance index.
[0048] Preferably, step S2 includes a weight allocation process:
[0049] The random forest algorithm was used to construct a feature importance assessment model, with the retained multidimensional features of drugs as input variables and the occurrence or non-occurrence of drug interactions as output variables to train the random forest model;
[0050] During the training process, the information gain and Gini index reduction of each feature in the decision tree construction process were calculated, and the importance score of each feature for predicting drug interactions was determined based on the calculated information gain and Gini index reduction. It should be further explained that in the specific implementation process, the random forest algorithm was used to evaluate the feature importance, which fully considered the complex interactive relationship between the features, and could more objectively reflect the contribution of each feature to the prediction of drug interactions, avoiding the subjectivity and one-sidedness of artificially setting weights.
[0051] It should be further explained that, in the specific implementation process, if a target feature greatly improves the node purity and has a higher information gain during the decision tree splitting process, the importance score of the target feature will be higher;
[0052] The weights are assigned according to the importance scores. The importance scores of each feature are converted into weight values by using the normalization method so that the sum of all weights is 1.
[0053] It should be further explained that, in the specific implementation process, if the importance score of a target feature is 0.3, its weight value after normalization is 0.3 / the sum of the importance scores of all features, ensuring that important features play a greater role in the prediction model and reducing false positive results caused by unreasonable feature weights.
[0054] Preferably, in step S3, the multiple drug interaction prediction models include a chemical structure characteristic model, a pharmacological action characteristic model, a metabolic pathway characteristic model and a gene expression characteristic model.
[0055] Preferably, step S3 includes constructing multiple drug interaction prediction models, the process is as follows:
[0056] For the chemical structure feature model, the chemical structure features of the drug are obtained, the chemical structure network is constructed, and the influence of the chemical structure features on drug interactions is analyzed through molecular dynamics simulation to obtain the structure prediction results;
[0057] For the pharmacological action characteristic model, the pharmacological action of the drug is obtained, the drug target network is constructed, and the network pharmacology is used to analyze how the drug exerts its effects by affecting multiple targets and biological pathways, as well as the potential interactions with the drug, to obtain the pharmacological prediction results;
[0058] For drug metabolism pathway models, we collect enzyme-catalyzed reactions and metabolites during drug metabolism in the body, build a metabolic pathway network, analyze the interaction relationship between drugs in the metabolic pathway, use pathway analysis algorithms to calculate the possibility of drug interactions, and obtain metabolic prediction results;
[0059] For the gene expression characteristic model, the change data of gene expression under the action of drugs is obtained, the gene regulatory network is constructed, and the network analysis method is used to predict the impact of drug interactions on gene expression and the drug interactions caused by them, and the gene prediction results are obtained.
[0060] Preferably, step S3 further includes: based on the chemical structure characteristic model, the pharmacological action characteristic model, the metabolic pathway characteristic model and the gene expression characteristic model, for each model, the collected data set is divided into 70% training set, 15% validation set and 15% test set, and the trained model is evaluated using the test set based on the evaluation indicators, and the evaluation indicators include accuracy, recall rate, F1 value and mean square error; the accuracy and reliability of each model in predicting drug interactions are measured by the evaluation indicators, and weight allocation is performed based on the comparison results of the evaluation indicators of each model on the test set;
[0061] For new drug data to be predicted, they are input into the chemical structure characteristic model, pharmacological action characteristic model, metabolic pathway characteristic model and gene expression characteristic model respectively to obtain the prediction results of each model.
[0062] Normalized weights are assigned according to feature importance scores so that each feature plays a role that matches its importance in the model, highlighting the influence of important features, optimizing the model's decision-making process, further improving prediction accuracy, and reducing the probability of false positive results.
[0063] Preferably, for obtaining the prediction results of each model, the prediction results include binary results or probability values; the binary results indicate whether there is an interaction, and the probability values indicate the probability of the interaction occurring when the drugs have an interaction;
[0064] If the predictions were binary, the predictions of each model were converted to numerical values: 1 for interaction and 0 for no interaction;
[0065] Then, a weighted average is performed according to the determined weights, and the judgment threshold is set to 0.5. If the weighted average value is greater than the judgment threshold, the final prediction result is that there is interaction, otherwise there is no interaction;
[0066] If the prediction result is a probability value, directly multiply the prediction probability of each model by the corresponding weight and add them together to get the final prediction probability value, which is used as the final prediction result;
[0067] It should be further explained that, in the specific implementation process, the chemical structure characteristic model prediction result is 1, the pharmacological action characteristic model prediction result is 0, the metabolic pathway characteristic model prediction result is 1, and the gene expression characteristic model prediction result is 1. The weighted average value is calculated according to the above weights: 0.3×1+0.25×0+0.2×1+0.25×1=0.8.
[0068] Preferably, step S1 includes:
[0069] For the multidimensional characteristics of drugs, the Pearson correlation coefficient of the occurrence and severity of the interaction between clinical data and drug characteristics is calculated to obtain the significance level and Pearson correlation coefficient; according to the size and significance level of the Pearson correlation coefficient, the characteristics with strong correlation are screened to obtain the multidimensional characteristics of drugs;
[0070] The significance level was set at 0.05. If the p-value corresponding to the Pearson correlation coefficient of the drug feature was less than 0.05 and the absolute value of the Pearson correlation coefficient was greater than 0.3, the feature was considered to be closely related to the drug interaction and was retained. Otherwise, if the p-value was greater than 0.05 or the absolute value of the Pearson correlation coefficient was less than 0.3, the feature was removed.
[0071] It should be further explained that, in the specific implementation process, feature correlation analysis based on a large amount of real clinical case data can more accurately identify features that are closely related to the actual occurrence of drug interactions, avoid feature misselection due to theoretical assumptions or limited data, and improve the reliability and effectiveness of feature screening;
[0072] Through rigorous statistical tests and correlation coefficient evaluations, we can scientifically determine whether to retain or remove features, reduce irrelevant or interfering features entering the model, focus the model on the key factors that truly affect drug interactions, and improve the model's predictive performance.
[0073] Obtain the network topology information of drug interactions and construct a drug interaction network; Based on the drug interaction network, combine the neighborhood overlap similarity algorithm and the Pearson correlation coefficient to train the graph convolutional neural network model;
[0074] In the message passing stage, the neighborhood overlap similarity between the drug center node and its adjacent nodes is calculated, and the weight coefficient is obtained through data processing. At the same time, the intermediate feature matrix is constructed and regularized to obtain the drug node feature matrix. The weight coefficient, edge set and drug node feature matrix are input into the model training.
[0075] In the message aggregation stage, the Pearson correlation coefficient between the feature vector of the central node and its adjacent nodes in the feature space is calculated to influence the feature weights of the adjacent nodes;
[0076] The process of building and training the graph convolutional neural network model is as follows:
[0077] Based on the drug interaction network, the node feature vectors are combined with the network topology information to construct a graph convolutional neural network model, which is denoted as the GCN model. In the message passing stage of the graph convolutional neural network model, the neighborhood overlap similarity between the drug center node and its adjacent nodes is calculated.
[0078] The formula is: ;in It's medicine The first-order neighborhood set in G, It's medicine the set of first-order neighbors in G;
[0079] in, : Indicates medicine With drugs Neighborhood overlap similarity between ; , : Represent two different drugs; :drug The first-order neighborhood set in the drug interaction network G, that is, the drug The set of directly connected drug nodes; :drug The first-order neighborhood set in the drug interaction network G, that is, the drug The set of directly connected drug nodes; :express and The number of elements in the intersection of and medications The number of common first-order neighbor nodes;
[0080] After obtaining the neighborhood overlap similarity, input the weight function: for processing, and then use a layer of perceptron to learn the weight function, and after the dropout layer, output the final weight coefficient , where the random probability of the dropout layer is set to 0.5;
[0081] Among them, the weight function formula is:
[0082] ;
[0083] in, :It is an overlap similarity with the neighborhood The associated weight function value vector; is a preset parameter used to adjust the sensitivity of the weight function to the neighborhood overlap similarity; is a natural constant;
[0084] At the same time, an intermediate feature matrix is constructed based on the node feature vector, and regularization is performed on it to obtain the drug node feature matrix; the weight coefficient, edge set and drug node feature matrix are input into the GCN model for training;
[0085] In the message aggregation stage, the Pearson correlation coefficient between the feature vector of the central node and its adjacent node feature vectors in the feature space is calculated:
[0086] The formula is:
[0087] ;
[0088] in, : Indicates a variable With variables Pearson correlation coefficient between ; , : Here, they represent the feature vector of the drug center node and its adjacent node feature vectors respectively; : represents the number of samples. When calculating the correlation of drug feature vectors, it can be understood as the number of data points used to calculate the correlation coefficient; , :Represents variables respectively and The data points; , :Represents variables respectively and The mean of
[0089] Adjust the feature weights of adjacent nodes according to the Pearson correlation coefficient, and output the final representation vector of each node, which is recorded as the prediction result;
[0090] Negative sampling technology was used to obtain the same number of non-existent edges as those in the drug interaction network as negative samples, and a sample data set was constructed together with the existing edges, where the existing edges were used as positive samples; the data set was divided into training set, validation set and test set at a ratio of 7:2:1 by the holdout method; the GCN model was trained using the training set, and after each round of training, the model prediction value was compared with the true value, and the binary classification model evaluation index was calculated on the validation set, and the model parameters were adjusted according to the evaluation index. The model with the highest evaluation index value in all rounds of training was selected as the optimal GCN model;
[0091] It should be further explained that in the specific implementation process, when the accuracy of the GCN model on the validation set is 0.9, the accuracy of the drug metabolism pathway model is 0.85, and the accuracy of the gene expression feature model is 0.8, the GCN model is assigned a weight of 0.9 / (0.9+0.85+0.8) according to the accuracy ratio, and so on; then the final integrated prediction result is obtained by comprehensive calculation.
[0092] It should be further explained that, in the specific implementation process, through integrated learning and model fusion, the complementary advantages of different models are utilized, and the reasonable screening and weighting of features can enable the model to maintain relatively stable performance in different data sets and application scenarios, reduce model prediction bias caused by data fluctuations or feature changes, enhance the generalization ability of the model, improve its applicability and reliability in the actual drug interaction prediction, and reduce the probability of a single model producing false positive results.
[0093] Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field and related fields without creative work should fall within the scope of protection of the present invention. The structures, devices and operating methods not specifically described and explained in the present invention are implemented according to the conventional means in the field unless otherwise specified and limited.
Claims
1. A drug interaction prediction method based on multidimensional features, characterized in that: The following steps are involved: S1: Collect multidimensional feature information of drugs from authoritative databases such as DrugBank, KEGG, and Pubchem to obtain drug features; collect real clinical case data, record them as clinical data, analyze the features that are closely related to the actual interaction between clinical data and drug features, remove drug features with weak correlation or prone to misjudgment, screen out features that are closely related to drug interactions, and obtain multidimensional drug features; S2: Based on the multidimensional features of drugs obtained after screening, the feature importance evaluation algorithm in machine learning is used to calculate the relative importance of each multidimensional feature of drugs in predicting drug interactions and obtain the importance index; According to the importance index of the multidimensional characteristics of the drug, a corresponding weight is assigned to each feature to obtain the weight distribution result; S3: Based on the multidimensional characteristics of drugs, multiple drug interaction prediction models are constructed, and the prediction results of multiple prediction models are integrated. The weighted average method is used to assign different weights to each prediction result according to the accuracy and reliability of each model to obtain the final prediction result; S4: Compare the final prediction result with the collected real-world drug use data to obtain a comparison result, and provide the final prediction result to the clinician based on the comparison result; Step S1 includes collaborating with medical institutions to collect real-world drug use data, which is recorded as real data. The real data includes actual reactions of patients after taking multiple drugs at the same time and adverse reaction monitoring data; In step S2, the relative importance of each drug multidimensional feature in predicting drug interactions is calculated to obtain an importance index. The process is as follows: Based on the multidimensional characteristics of the drugs obtained after screening, the drug interaction result data corresponding to the drug samples are prepared as labels to obtain binary labels or probability values. The binary labels include interaction and no interaction. The probability value indicates the probability of interaction when the drugs interact. Based on random forest as the feature importance evaluation algorithm, the contribution of each feature to reducing impurity when building a decision tree is calculated to evaluate the feature importance. In the case of multiple trees, the importance of the features in each tree is averaged to obtain the final feature importance evaluation result. Use drug multidimensional features and corresponding drug interaction label data to train a random forest model, build a random forest model, output the importance score of each feature, sort the calculated importance results of each feature from high to low according to the importance score, and obtain the importance sequence of each drug multidimensional feature in predicting drug interactions. Based on the importance sequence, combined with a preset importance index comparison list, match the position of the drug feature in the importance sequence with the importance index matched in the importance index comparison list; Step S2 includes the weight assignment process: The random forest algorithm was used to construct a feature importance assessment model, with the retained multidimensional features of drugs as input variables and the occurrence or non-occurrence of drug interactions as output variables to train the random forest model; During the training process, the information gain and Gini index reduction of each feature in the decision tree construction process were calculated, and the importance score of each feature for predicting drug interactions was determined based on the calculated information gain and Gini index reduction; The weights are assigned according to the importance scores. The importance scores of each feature are converted into weight values by using the normalization method so that the sum of all weights is 1. In step S3, the multiple drug interaction prediction models include a chemical structure characteristic model, a pharmacological action characteristic model, a metabolic pathway characteristic model, and a gene expression characteristic model; Step S3 includes constructing multiple drug interaction prediction models, and the process is as follows: For the chemical structure feature model, the chemical structure features of the drug are obtained, the chemical structure network is constructed, and the influence of the chemical structure features on drug interactions is analyzed through molecular dynamics simulation to obtain the structure prediction results; For the pharmacological action characteristic model, the pharmacological action of the drug is obtained, the drug target network is constructed, and the network pharmacology is used to analyze how the drug exerts its effects by affecting multiple targets and biological pathways, as well as the potential interactions with the drug, to obtain the pharmacological prediction results; For drug metabolism pathway models, we collect enzyme-catalyzed reactions and metabolites during drug metabolism in the body, build a metabolic pathway network, analyze the interaction relationship between drugs in the metabolic pathway, use pathway analysis algorithms to calculate the possibility of drug interactions, and obtain metabolic prediction results; For the gene expression feature model, obtain the data on changes in gene expression under the action of drugs, build a gene regulatory network, use network analysis methods to predict the effects of drug interactions on gene expression and the drug interactions caused by them, and obtain gene prediction results; Step S1 includes: For the multidimensional characteristics of drugs, the Pearson correlation coefficient of the occurrence and severity of the interaction between clinical data and drug characteristics is calculated to obtain the significance level and Pearson correlation coefficient; according to the size and significance level of the Pearson correlation coefficient, the characteristics with strong correlation are screened to obtain the multidimensional characteristics of drugs; The significance level is set to A. If the p-value corresponding to the Pearson correlation coefficient of the drug feature is less than A and the absolute value of the Pearson correlation coefficient is greater than B, the feature is considered to be closely related to the drug interaction and is retained; otherwise, if the p-value is greater than A or the absolute value of the Pearson correlation coefficient is less than B, the feature is removed.
2. The method for predicting drug interactions based on multidimensional features according to claim 1, characterized in that: The multidimensional characteristics of the drug include the chemical structure characteristics, pharmacological action characteristics, metabolic pathway characteristics and gene expression characteristics of the drug.
3. The method for predicting drug interactions based on multidimensional features according to claim 2, characterized in that: Step S3 also includes: based on the chemical structure feature model, the pharmacological action feature model, the metabolic pathway feature model and the gene expression feature model, for each model, dividing the collected data set in proportion, and evaluating the trained model using the test set based on the evaluation indicators, the evaluation indicators including accuracy, recall rate, F1 value and mean square error; and weighting each model based on the comparison results of the evaluation indicators on the test set; For new drug data to be predicted, they are input into the chemical structure characteristic model, pharmacological action characteristic model, metabolic pathway characteristic model and gene expression characteristic model respectively to obtain the prediction results of each model.
4. The method for predicting drug interactions based on multidimensional features according to claim 3, characterized in that: The prediction results of each model include binary results or probability values; the binary results indicate whether there is an interaction or not, and the probability value indicates the probability of an interaction when the drugs have an interaction; If the predictions were binary, the predictions of each model were converted to numeric values: 1 for interaction and 0 for no interaction; Then, a weighted average is performed according to the determined weights, and the judgment threshold is set to 0.
5. If the weighted average is greater than the judgment threshold, the final prediction result is that there is interaction, otherwise there is no interaction; If the prediction result is a probability value, simply multiply the prediction probability of each model by the corresponding weight and add them together to obtain the final prediction probability value, which is used as the final prediction result.