Drug combination screening method based on multi-level interpretable graph convolutional network

Abstract

The invention belongs to the technical field of medicine, bioinformatics and computational biology, and particularly relates to a drug combination screening method based on a multi-level interpretable graph convolutional network, and the method comprises the steps: obtaining an effective region of a drug to a disease according to the correlation between the drug and a disease target; acquiring a network distance between the disease and the drug target according to a network relationship between the disease and the drug target; obtaining medication experience; according to the chemical components of the medicine and the disease target, obtaining the type of the compound; inputting the effective area, the network distance, the medication experience and the compound category into a graph convolutional network model to obtain a correlation value of the medicine; based on the correlation degree value, obtaining a drug correlation coefficient; and according to the correlation coefficient, obtaining the pharmaceutical composition. According to the invention, through prediction of the graph convolution model, the drug combination for treating diseases can be accurately predicted.

Description

Technical Field

[0001] The present invention belongs to the multi-technical fields of medicine, bioinformatics, and computational biology, and particularly relates to a method for screening drug combinations based on a multi-level interpretable graph convolutional network. Background Art

[0002] Although monotherapy has made remarkable progress in the field of disease treatment, it still has limitations such as drug response heterogeneity, toxicity, and drug resistance, which have prompted multi-target and drug combination therapies to become an important strategy for treating complex diseases. Drug combinations can act on multiple targets, pathways, and mechanisms, improving efficacy, reducing side effects, and decreasing drug resistance. Although the traditional trial-and-error method is accurate and effective, it is time-consuming and laborious in drug combination design. Therefore, models based on high-throughput screening and artificial intelligence (AI) have emerged. These models optimize drug combination prediction by automatically learning complex data features and reduce the dependence on dose-response relationships. With the in-depth research and technological innovation, using deep learning (DL) models, researchers have adopted various strategies to optimize performance, such as regularization techniques and gradient boosting tree methods, and developed various models such as DeepSynergy to predict the synergistic effects of drug combinations, demonstrating the advantages of DL in handling non-linear relationships and complex data. However, these models still face challenges such as limited benchmark dataset size, incomplete target information, and insufficient adverse reaction prediction ability. Generally speaking, AI technology provides strong support for the research and application of drug combinations, significantly improving prediction accuracy and efficiency.

[0003] With the development of big data, multi-omics technologies, and AI, the integration of traditional Chinese medicine and AI has become increasingly close, bringing new opportunities to the theory and practice of traditional Chinese medicine. Traditional Chinese medicine pays attention to the compatibility relationship between drugs, but there are limitations in explaining its modern molecular mechanisms and treatment principles. Summary of the Invention

[0004] To solve the above technical problems, the present invention proposes a method for screening drug combinations based on a multi-level interpretable graph convolutional network, which can screen out potential effective drugs and compounds for diseases.

[0005] The present invention provides a method for screening drug combinations based on a multi-level interpretable graph convolutional network, including:

[0006] Obtaining the effective region of a drug for a disease according to the association between the drug and the disease target;

[0007] Obtaining the network distance between the disease and the drug target according to the network relationship between the disease and the drug target;

[0008] Obtaining medication experience;

[0009] Obtain the compound category according to the chemical composition of the drug and the disease target;

[0010] Input the effective region, network distance, medication experience, and compound category into the graph convolutional network model to obtain the association degree value of the drug;

[0011] Based on the association degree value, obtain the drug correlation coefficient;

[0012] Obtain the drug combination according to the correlation coefficient;

[0013] Optionally, obtaining the effective region of the drug for the disease according to the association between the drug and the disease target includes:

[0014] Obtain the overlap rate and H score of the drug for the disease according to the association between the drug and the disease target;

[0015] Obtain the effective region of the drug for the disease according to the overlap rate and H score;

[0016] Optionally, obtaining the network distance between the disease and the drug target according to the network relationship between the disease and the drug target includes:

[0017] Obtain the network relationship between the disease and the drug target according to the similarity of the network distances of the disease genes;

[0018] Obtain the average shortest distance within the group and the average shortest distance between the disease-drug target pairs according to the network relationship between the disease and the drug target;

[0019] Based on the average shortest distance within the group and the average shortest distance between the disease-drug target pairs, obtain the network distance between the disease and the drug target;

[0020] Optionally, obtaining the medication experience includes:

[0021] Obtain the drug pairs based on the database;

[0022] Calculate the Pscore value and Mscore value of the corresponding drug pairs, and obtain the medication experience according to the Pscore value and Mscore value;

[0023] Optionally, obtaining the compound category according to the chemical composition of the drug and the disease target includes:

[0024] Obtain the chemical composition of the drug, perform molecular docking on the chemical components whose occurrence frequency reaches the preset number with the disease target, and obtain the molecular fingerprint map after docking;

[0025] Based on the molecular fingerprint map, obtain the molecular fingerprint similarity;

[0026] Classify the molecular fingerprint similarity to obtain the compound category.

[0027] Optionally, the graph convolutional network model includes: an input layer, a GCN layer, a prediction layer, and an output layer;

[0028] The input layer is used to input the effective region, network distance, medication experience, and compound category;

[0029] The GCN layer is used to aggregate the pre-processed data and update the feature representation of the nodes or edges in the graph, that is, construct a matrix to capture the relationship between the structural information of the graph and the node features;

[0030] The prediction layer is used to predict the correlation value of the drug;

[0031] The output layer is used to output the correlation value of the drug.

[0032] Optionally, based on the correlation value, obtaining the drug correlation coefficient includes:

[0033] Based on the correlation value, use the singular value decomposition method to obtain the drug correlation coefficient.

[0034] Optionally, using the singular value decomposition method to obtain the drug correlation coefficient includes:

[0035] Perform singular value decomposition on the matrix to obtain the corresponding principal components;

[0036] Calculate the correlation coefficient between the corresponding principal components by vector inner product to obtain the drug correlation coefficient.

[0037] Compared with the prior art, the present invention has the following advantages and technical effects:

[0038] The present invention filters and screens through the target repeatability of the drug, the network topology distance correlation, and the clinical medication experience, and can determine a set of non-repetitive drug groups included in a certain type of disease treatment; the present invention also constructs an effective chemical component set for treating diseases as the background data set; the present invention divides all effective compounds into a certain number of spaces, and the closer the spatial distance of structurally similar compounds, and can divide structurally similar drugs into the same drug space; through the prediction of the graph convolutional model, the present invention can more accurately predict the drug combinations for treating diseases. Description of the Drawings

[0039] The drawings constituting a part of this application are used to provide a further understanding of this application. The schematic embodiments of this application and their descriptions are used to explain this application and do not constitute an improper limitation to this application. In the drawings:

[0040] Figure 1It is a flowchart of a method for screening drug combinations based on a multi-level interpretable graph convolutional network according to an embodiment of the present invention;

[0041] Figure 2 It is a framework diagram of a method for screening drug combinations based on a multi-level interpretable graph convolutional network according to an embodiment of the present invention. Detailed implementation manners

[0042] It should be noted that, without conflict, the embodiments in the present application and the features in the embodiments may be combined with each other. The present application will be described in detail below with reference to the accompanying drawings and in conjunction with the embodiments.

[0043] It should be noted that the steps shown in the flowchart of the accompanying drawings may be executed in a computer system such as a set of computer-executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described may be executed in a different order from that here.

[0044] This embodiment proposes a method for screening drug combinations based on a multi-level interpretable graph convolutional network, as Figure 1 shown, and specifically includes the following steps:

[0045] According to the association between drugs and disease targets, obtain the effective area of the drug for the disease;

[0046] According to the network relationship between the disease and the drug target, obtain the network distance between the disease and the drug target;

[0047] Obtain medication experience;

[0048] According to the chemical composition of the drug and the disease target, obtain the compound category;

[0049] Input the effective area, network distance, medication experience, and compound category into the graph convolutional network model to obtain the association degree value of the drug;

[0050] Based on the association degree value, obtain the drug correlation coefficient;

[0051] According to the correlation coefficient, obtain the drug combination.

[0052] Specifically, based on the area overlap rate, PageRank, and two-step random walk algorithm, obtain the association between a single drug and a disease target, and obtain the effectiveness Overlap rate and Hscore of a certain drug for the disease;

[0053] Based on the similarity of the network distance of disease genes, calculate the network relationship between the disease-drug target, and obtain the average shortest distance s within the group AB and the average shortest distance d between disease-drug target pairs AB ;

[0054] By retrieving, de-duplicating, and sorting the drug pairs in the database of traditional Chinese medicine value evaluation information, the Pscore value and Mscore value of the corresponding drug pairs are obtained;

[0055] Summarize the chemical components of the drugs, and perform molecular docking on the compounds with a frequency of occurrence ≥ 5 times and the disease targets. Based on the principle of measuring the molecular fingerprint similarity by the Tanimoto coefficient, use the K-means clustering method to classify the compounds within the drugs;

[0056] Construct a graph convolutional network model, including the graph structure data of disease genes and drug nodes, and calculate the importance of the edges and the correlation degree values between drugs. Based on the singular value decomposition method, calculate the row ranking of the average correlation coefficient of drugs, and select potential drug combinations for treating diseases.

[0057] Furthermore, according to the association between the drugs and the disease targets, the effective regions of the drugs for the diseases are obtained, including:

[0058] According to the association between the drugs and the disease targets, obtain the overlap rate and H score of the drugs for the diseases;

[0059] According to the overlap rate and H score, obtain the effective regions of the drugs for the diseases.

[0060] Specifically, connect the drug-compound-disease network through the PageRank and two-step random walk algorithms. Among them, the drug contains more components, and these components have a greater impact on the target. The more disease targets the drug acts on, the more important these targets are in the PPI network, and the more effective the drug is in treating the disease, and the higher the Hscore. The specific calculation formula is as follows:

[0061]

[0062] In the formula, the Overlap rate is the area overlap rate, represents the number of genes of the drug, n Hub-D represents the number of genes of the disease, represents the number of genes in the intersection of the drug and disease genes.

[0063]

[0064] In the formula, Hscoreγ i is the effectiveness score of a certain drug for the disease, A β represents the targeting score for β, V βγ then represents the matrix constructed between β and γ, where β represents the degree value of a certain drug, γ is the degree value of the disease, and 1 and 0 represent whether there is a connection between the two, β out then represents the out-degree value of the drug node β in the network.

[0065] Further, obtaining the network distance between a disease and a drug target according to the network relationship between the disease and the drug target includes:

[0066] Obtaining the network relationship between the disease and the drug target according to the similarity of the network distances of the disease genes; obtaining the average shortest distance within the group and the average shortest distance between the disease-drug target pairs according to the network relationship between the disease and the drug target, where drugs A and B and the disease are grouped together;

[0067] Obtaining the network distance between the disease and the drug target based on the average shortest distance within the group and the average shortest distance between the disease-drug target pairs.

[0068] Specifically, based on the genes corresponding to the drugs forming modules in the protein interactome, using the similarity of their network distances to the disease genes as background data, calculating the network relationship between the disease-drug targets, and obtaining the average shortest distance s within the group AB and the average shortest distance d between the disease-drug target pairs AB ;

[0069]

[0070] In the formula, d AA and d BB represent the average shortest distance within the interaction region between the two nodes of A and B. When s AB < 0, the two groups of targets are topologically overlapping; when s AB ≥ 0, the two groups of targets are topologically separated.

[0071] The Z value is a reliable indicator for measuring the network proximity between a drug target (X) and a disease (Y). The shortest path length d(x, y) between the drug target (X) and the disease (Y) can be calculated according to the following formula:

[0072]

[0073] In the formula, the mean is set as μ and the standard deviation is set as σ. From a network perspective, if the drug target and the disease target are separated from each other, then the corresponding z ≥ 0; otherwise z < 0.

[0074] Further, obtaining the medication experience includes:

[0075] Obtaining drug pairs based on a database;

[0076] Calculating the Pscore value and the Mscore value of the corresponding drug pairs, and obtaining the medication experience according to the Pscore value and the Mscore value.

[0077] Specifically, by retrieving, removing duplicates, and organizing 28,279 pairs of medicinal herb pairs in the database for the value evaluation of traditional Chinese patent medicines, the Pscore values of the corresponding medicinal herb pairs were calculated. All the drug compounds related to clinical diagnosis and treatment diseases retrieved from CNKI in the past 10 years were added. After pre-removing duplicates and normalizing, factor analysis was performed, and Mscore represents the average number of times a certain drug appears in common medicinal herb pairs:

[0078] Pscore = count(Herb i ∩Herb j )

[0079]

[0080] In the formula, both Herb i and Herb j represent drugs. Pscore j represents the sum of the Pscore values of all the medicinal herb pairs retrieved in the database that contain a certain drug Herb j . ε represents the maximum absolute value factor weight of a certain drug Herb i after factor analysis, and n j represents the number of the retrieved medicinal herb pairs.

[0081] Furthermore, according to the chemical components and disease targets of the drugs, the compound categories obtained include:

[0082] The chemical components of the drugs were obtained, and the chemical components that reached the preset number of occurrences were docked with the disease targets to obtain the molecular fingerprint map after docking;

[0083] Based on the molecular fingerprint map, the molecular fingerprint similarity was obtained;

[0084] The molecular fingerprint similarity was classified to obtain the compound categories.

[0085] Specifically, the compounds with the occurrence frequency ≥5 times were docked with the disease genes. The Pymol software was used to dehydrate and remove ligands from the receptor protein, the AutoDockTools software was used to hydrogenate and balance the charges of the receptor protein, and the AutoDock Vina was used for molecular docking scoring. The results were output as Affinity, which is an important indicator to measure whether the ligand can effectively bind to the receptor molecule. The "rcdk", "ChemmineR", and "cluster" packages were used to obtain the molecular fingerprint maps of the effective compounds, the Tanimoto coefficient was used to measure the molecular fingerprint similarity, and the K-means clustering method was used to classify the compounds in the drug into three categories according to their distances.

[0086] Furthermore, the graph convolutional network model includes: an input layer, a GCN layer, a prediction layer, and an output layer;

[0087] An input layer for inputting the effective region, network distance, medication experience, and compound category;

[0088] A GCN layer for aggregating the pre - processed data and updating the feature representation of nodes (or edges) in the graph, that is, constructing a matrix to capture the relationship between the structural information of the graph and the node features;

[0089] A prediction layer for predicting the association degree value of the drug;

[0090] An output layer for outputting the association degree value of the drug.

[0091] Specifically, the graph data used consists of disease - gene nodes and drug nodes, where each node represents a gene and the edges represent the interactions between genes. To construct such a graph structure, exemplary graph data is first generated, which contains randomly generated node features and edge indices. In addition, a binary label (0 or 1) is assigned to each edge to indicate whether the edge exists. To enhance the interpretability of the model, for each edge, the inner - product operation is performed using the embeddings of its two - end nodes to obtain a scalar as the importance score of the edge. Then, the score is mapped to the [0, 1] interval through the Sigmoid function to represent the probability of the edge existing. The model is trained using the Adam optimizer and the binary cross - entropy loss function. At the beginning of each training epoch, all parameter gradients are cleared, and then the probability distribution of the edges is calculated based on the current input, and the loss is calculated based on this. After the loss is back - propagated, the model parameters are updated. During the training process, gradient clipping technology is added to limit the maximum norm not to exceed 1.0, and the specific formula is as follows:

[0092]

[0093] H (l+1) =(AH (l) ·G (l) )

[0094]

[0095] In the formula, E C is the docking ability of the compound and the average effective coverage rate score of traditional Chinese medicine compounds, N is the number of drugs, N Herb represents the total number of effective compounds found after molecular docking, n Herb(j) represents the total number of compounds contained in the drug Herb j , represents the total number of effective compounds contained in the drug Herb j . A is the normalized adjacency matrix, H represents the attention feature matrix, G (l) is the weight matrix of the (l) - th layer. Score (Herb(i),Herb(j)) represents Herbj and Herb i The association score of the node pair, diag(w) represents converting the vector w into a diagonal matrix, and w represents a learnable dimensional parameter. Prob (Herb(i),Herb(j)) represents the drug Herb j and Herb i The probability of a synergistic effect between them, where σ represents the Sigmoid function. represents the drug Herb j and Herb i The association probability between them.

[0096] Furthermore, based on the association degree value, obtaining the drug correlation coefficient includes:

[0097] Based on the association degree value, using the singular value decomposition method to obtain the drug correlation coefficient.

[0098] Furthermore, using the singular value decomposition method to obtain the drug correlation coefficient includes:

[0099] Performing singular value decomposition on the matrix to obtain the corresponding principal components;

[0100] Calculating the correlation coefficient between the corresponding principal components by the inner product of vectors to obtain the drug correlation coefficient.

[0101] Specifically, singular value decomposition is a common linear transformation method widely used in statistics and machine learning. After information normalization, singular value decomposition is performed on the matrix to obtain the corresponding principal components, and the correlation coefficient between drugs is calculated by the inner product of vectors. This method innovatively considers the control ability of the characteristics of drugs themselves over diseases and compound nodes. The drugs predicted by the graph convolutional network model are ranked based on the correlation coefficient between pairwise drugs under singular value decomposition, and the degree value and the average drug correlation coefficient are used to select the drug combination for treating diseases.

[0102] The following combines the attached Figure 2 to elaborate on this embodiment in detail:

[0103] Taking the drug screening for ankylosing spondylitis as an example, the specific process includes:

[0104] A set of non-repeating drug groups for treating a certain type of disease is determined. Each traditional Chinese medicine is comprehensively scored through overlap rate calculation and Hscore. A total of 109 traditional Chinese medicines are screened out with the thresholds of Overlap rate > 0 and Hscore > 0.001. The PPI network is constructed for the targets of each traditional Chinese medicine, and the network proximity from the traditional Chinese medicine targets to the disease module is calculated respectively, and it is found that there is topological intersection (Z < 0) between the traditional Chinese medicine targets and the hub genes. A total of 105 traditional Chinese medicines are screened out with non-steroidal anti-inflammatory drugs as the threshold. Combining the experience of using traditional Chinese medicines, 45 traditional Chinese medicines are further screened out with Mscore > 5, which contain a total of 612 compounds.

[0105] The set of effective chemical components for treating the disease is used as the background data set. Combining the overlap between the genes corresponding to the traditional Chinese medicines and ankylosing spondylitis and the results of molecular docking, a total of 14 effective compounds are found.

[0106] Calculate the space of similar compounds and drug groups. The hierarchical and K-means clustering methods are used to further analyze 32 traditional Chinese medicines with the number of effective compounds ≥ 3, and a total of 24 traditional Chinese medicines are included for subsequent prediction by the graph convolutional network model. Each compound and drug space includes background data and traditional Chinese medicine chemical composition data.

[0107] The graph convolutional model predicts the drug combinations for treating ankylosing spondylitis. The parameters of the GCN model are pre-debugged, and the final running parameters are set as: hidden_dim = 32, output_dim = 1, learning_rate = 0.0001, epochs = 100. The training curve shows a state of first decreasing and then tending to be stable, which conforms to the expected optimization trend. There are still small-scale rebounds in the loss value in individual cycles, which may be caused by data complexity or insufficiently smooth gradient updates. A total of 225 traditional Chinese medicine prediction results are obtained in the final run, among which there are 146 entries with potential synergistic relationships between the drugs. By calculating the average correlation of each traditional Chinese medicine under SVD, with the correlation value > 10 and the correlation coefficient between the drugs > 0.5, as shown in Table 1, 8 traditional Chinese medicines are finally obtained: myrrh, drynaria rhizome, wolfberry fruit, epimedium, achyranthes root, galangal, forsythia, and astragalus membranaceus.

[0108] Table 1

[0109]

[0110] As shown in Table 2, it was compared with seven common ML models using a 3-fold cross-validation (CV) strategy, namely decision tree, random forest, gradient boosting tree, support vector machine, extreme gradient boosting, light gradient boosting machine, and k-nearest neighbor algorithm. The main outcome measures used were classification metrics accuracy, recall, precision, F1-score, area under the ROC curve, and regression metrics root mean square error, mean absolute error to compare the predicted values with the true values. The data was split with a 7:3 training to test set ratio.

[0111] Table 2

[0112]

[0113]

[0114] The experimental results show that compared with multiple methods including decision tree, random forest, gradient boosting tree, support vector machine, extreme gradient boosting, light gradient boosting machine, and k-nearest neighbor algorithm, this method demonstrates excellent performance in classification. Especially in key evaluation metrics such as accuracy, recall, precision, and F1-score, it significantly outperforms other models. Although this method did not reach the lowest values in regression tasks such as root mean square error and mean absolute error metrics, it was optimized for the unique properties of each drug and its clinical applications, showing comprehensive advantages. In summary, in the context of considering drug specificity and clinical usage, this method can not only effectively handle classification problems but also provide a more comprehensive solution in regression analysis.

[0115] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any changes or substitutions that can be easily thought of by those skilled in the art within the technical scope disclosed in the present application should be covered by the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A method for screening drug combinations based on a multi-level interpretable graph convolutional network, characterized in that, Including: Obtain the effective region of the drug against the disease according to the association between the drug and the disease target; Obtain the network distance between the disease and the drug target according to the network relationship between the disease and the drug target; Obtain the medication experience; Obtain the compound category according to the chemical composition of the drug and the disease target; Input the effective region, network distance, medication experience and compound category into the graph convolutional network model to obtain the association degree value of the drug; Based on the association degree value, obtain the drug correlation coefficient; Obtain the drug combination according to the correlation coefficient.

2. The method for screening drug combinations based on a multi-level interpretable graph convolutional network according to claim 1, wherein, Obtaining the effective region of the drug against the disease according to the association between the drug and the disease target includes: Obtain the overlap rate and H score of the drug against the disease according to the association between the drug and the disease target; Obtain the effective region of the drug against the disease according to the overlap rate and H score.

3. The method for screening drug combinations based on a multi-level interpretable graph convolutional network according to claim 1, wherein Obtaining the network distance between the disease and the drug target according to the network relationship between the disease and the drug target includes: Obtain the network relationship between the disease and the drug target according to the similarity of the network distance of the disease gene; According to the network relationship between the disease and the drug target, obtain the average shortest distance within the group and the average shortest distance between the disease-drug target pairs; Based on the average shortest distance within the group and the average shortest distance between the disease-drug target pairs, obtain the network distance between the disease and the drug target.

4. A method for screening drug combinations based on a multi-level interpretable graph convolutional network according to claim 1, characterized in that, Obtaining the medication experience includes: Based on the database, obtain the drug pairs; Calculate the Pscore value and Mscore value of the corresponding drug pairs, and obtain the medication experience according to the Pscore value and Mscore value.

5. A method for screening drug combinations based on a multi-level interpretable graph convolutional network according to claim 1, characterized in that Obtaining the compound category according to the chemical composition of the drug and the disease target includes: Obtain the chemical composition of the drug, perform molecular docking on the chemical components whose occurrence frequency reaches the preset number with the disease target, and obtain the molecular fingerprint map after docking; Based on the molecular fingerprint map, obtain the molecular fingerprint similarity; Classify the molecular fingerprint similarity to obtain the compound category.

6. The method for screening drug combinations based on a multi-level interpretable graph convolutional network according to claim 1, wherein, The graph convolutional network model includes: an input layer, a GCN layer, a prediction layer and an output layer; The input layer is used to input the effective region, network distance, medication experience and compound category; The GCN layer is used to aggregate the pre-processed data in the early stage and update the feature representation of the nodes or edges in the graph, that is, construct a matrix to capture the relationship between the structural information of the graph and the node features; The prediction layer is used to predict the association degree value of the drug; The output layer is used to output the association degree value of the drug.

7. A method for screening drug combinations based on a multi-level interpretable graph convolutional network according to claim 6, characterized in that Based on the association degree value, obtaining the drug correlation coefficient includes: Based on the association degree value, use the singular value decomposition method to obtain the drug correlation coefficient.

8. A method for screening drug combinations based on a multi-level interpretable graph convolutional network according to claim 7, characterized in that Using the singular value decomposition method to obtain the drug correlation coefficient includes: Perform singular value decomposition on the matrix to obtain the corresponding principal components; Calculate the correlation coefficient between the corresponding principal components by vector inner product to obtain the drug correlation coefficient.

Images