A method for drug combination screening based on a multi-level interpretable graph convolutional network

Abstract

The application belongs to the technical fields of medicine, bioinformatics and computational biology, and particularly relates to a drug combination screening method based on a multi-level interpretable graph convolution network, which comprises the following steps: obtaining an effective area of a drug on a disease according to the correlation between the drug and a disease target; obtaining a network distance of the disease and the drug target according to the network relationship between the disease and the drug target; obtaining drug experience; obtaining a compound category according to the chemical composition of the drug and the disease target; inputting the effective area, the network distance, the drug experience and the compound category into a graph convolution network model to obtain a correlation degree value of the drug; obtaining a drug correlation coefficient based on the correlation degree value; and obtaining a drug combination according to the correlation coefficient. Through the prediction of the graph convolution model, the drug combination for treating the disease can be accurately predicted.

Description

TECHNICAL FIELD

[0001] The present application belongs to the multi-technical field of medicine, bioinformatics and computational biology, and particularly relates to a drug combination screening method based on a multi-level interpretable graph convolution network. BACKGROUND

[0002] Although single-drug therapy has made significant progress in the field of disease treatment, it still has limitations such as drug response heterogeneity, toxicity and drug resistance, prompting multi-target and drug combination therapy to become an important strategy for the treatment of complex diseases. Drug combination can enhance efficacy, reduce side effects and reduce drug resistance by acting on multiple targets, pathways and mechanisms. Although the traditional trial-and-error method is accurate and effective, it is time-consuming and laborious in drug combination design, so models based on high-throughput screening and artificial intelligence (AI) have emerged. These models optimize drug combination prediction by automatically learning complex data features, reducing the dependence on dose-response relationships. With the in-depth research and technological innovation, researchers use various strategies to optimize performance, such as regularization techniques and gradient boosting tree methods, and develop models such as DeepSynergy to predict the synergistic effect of drug combinations, demonstrating the advantages of DL in handling nonlinear relationships and complex data. However, these models still face challenges such as limited size of benchmark data sets, incomplete target information, and insufficient adverse reaction prediction capabilities. Overall, 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 technology and AI, the integration of traditional Chinese medicine and AI is becoming increasingly close, bringing new opportunities for the theory and practice of traditional Chinese medicine. Traditional Chinese medicine focuses on the compatibility relationship between drugs, but there are limitations in explaining its modern molecular mechanisms and treatment principles. SUMMARY

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

[0005] The present application provides a drug combination screening method based on a multi-level interpretable graph convolution network, comprising:

[0006] According to the association between drugs and disease targets, the effective area of drugs on diseases is obtained;

[0007] According to the network relationship between diseases and drug targets, the network distance between diseases and drug targets is obtained;

[0008] Obtain drug experience;

[0009] Based on the chemical composition of the drug and the disease target, obtain the compound category;

[0010] The effective region, network distance, medication experience, and compound category are input into the graph convolutional network model to obtain the drug correlation value;

[0011] Based on the correlation value, the drug correlation coefficient is obtained;

[0012] Based on the correlation coefficient, drug combinations are obtained.

[0013] Optionally, based on the association between the drug and the disease target, the effective region of the drug for the disease can be identified, including:

[0014] Based on the association between drugs and disease targets, obtain the drug-disease overlap rate and H score;

[0015] Based on the overlap rate and H score, the effective region of the drug for the disease is obtained.

[0016] Optionally, based on the network relationship between the disease and the drug target, the network distance between the disease and the drug target can be obtained, including:

[0017] Based on the similarity of the network distance of disease genes, the network relationship between diseases and drug targets can be obtained;

[0018] Based on the network relationship between the disease and drug target, obtain the average shortest distance within the group and the average shortest distance between disease-drug target pairs;

[0019] The network distance between the disease and the drug target is obtained based on the average shortest distance within the group and the average shortest distance between the disease-drug target pairs.

[0020] Optional, gaining experience with medication includes:

[0021] Based on the database, obtain drug pairs;

[0022] Calculate the Pscore and Mscore values ​​for the corresponding drug pairs, and obtain the medication experience based on the Pscore and Mscore values.

[0023] Optionally, based on the drug's chemical composition and disease target, the compound categories obtained include:

[0024] Obtain the chemical components of the drug, perform molecular docking between chemical components that appear at a preset frequency and disease targets, and obtain the molecular fingerprint spectrum after docking.

[0025] Based on the molecular fingerprint spectrum, the molecular fingerprint similarity is obtained;

[0026] The molecular fingerprint similarity is classified 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 area, network distance, medication experience, and compound category;

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

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

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

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

[0033] Based on the correlation value, the drug correlation coefficient is obtained using singular value decomposition.

[0034] Optionally, obtaining the drug correlation coefficient using singular value decomposition includes:

[0035] Singular value decomposition is performed on the matrix to obtain the corresponding principal components;

[0036] The correlation coefficients between the corresponding principal components are calculated using the vector inner product to obtain the drug correlation coefficients.

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

[0038] This invention filters and selects drugs based on target repetition, network topology distance correlation, and clinical medication experience to identify a set of non-repeating drug groups for treating a particular disease. It also constructs a set of effective chemical components for treating diseases as background data. Furthermore, this invention divides all effective compounds into a certain number of spaces; compounds with similar structures are more closely spaced, allowing for the grouping of structurally similar drugs into the same drug space. Finally, this invention uses a graph convolution model to predict drug combinations for treating diseases with relatively high accuracy. Attached Figure Description

[0039] The accompanying drawings, which form part of this application, are used to provide a further understanding of this application. The illustrative embodiments and descriptions of this application are used to explain this application and do not constitute an undue limitation of this application. In the drawings:

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

[0041] Figure 2 This is a framework diagram of a drug combination screening method based on a multi-level interpretable graph convolutional network according to an embodiment of the present invention. Detailed Implementation

[0042] It should be noted that, unless otherwise specified, the embodiments and features described in this application can be combined with each other. This application will now be described in detail with reference to the accompanying drawings and embodiments.

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

[0044] This embodiment proposes a method for drug combination screening based on multi-level interpretable graph convolutional networks, such as... Figure 1 As shown, the specific steps include:

[0045] Based on the correlation between drugs and disease targets, the effective area of ​​the drug for the disease can be identified.

[0046] Based on the network relationship between diseases and drug targets, obtain the network distance between diseases and drug targets;

[0047] Gain experience in medication use;

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

[0049] The effective region, network distance, medication experience, and compound category are input into the graph convolutional network model to obtain the drug correlation value;

[0050] Based on the correlation score, obtain the drug correlation coefficient;

[0051] Based on the correlation coefficient, obtain the drug combination.

[0052] Specifically, based on the area overlap rate and PageRank and two-step random walk algorithms, the association between a single drug and a disease target is obtained, and the effectiveness of a certain drug against the disease is obtained by the overlap rate and Hscore.

[0053] Based on the similarity of network distances between disease genes, the network relationship between diseases and drug targets is calculated to obtain the average shortest distance s within a group. AB The average shortest distance d between the disease-drug target pair AB ;

[0054] By retrieving, deduplicating, and organizing drug pairs from the database of traditional Chinese medicine value assessment information, the Pscore and Mscore values ​​of the corresponding drug pairs are obtained.

[0055] The chemical composition of the drugs was summarized, and molecular docking was performed on compounds that appeared ≥5 times with disease targets. Based on the principle of measuring molecular fingerprint similarity using the Tanimoto coefficient, K-means clustering was used to classify the compounds within the drugs.

[0056] A graph convolutional network model is constructed, which includes graph structure data of disease genes and drug nodes. The importance of edges and the correlation between drugs are calculated. Based on the singular value decomposition method, the average correlation coefficient of drugs is calculated and the row ranking is determined to select potential drug combinations for treating diseases.

[0057] Furthermore, based on the correlation between drugs and disease targets, the effective regions of drugs for disease can be identified, including:

[0058] Based on the association between drugs and disease targets, obtain the drug-disease overlap rate and H score;

[0059] Based on the overlap rate and H score, the effective region of the drug for the disease is obtained.

[0060] Specifically, the drug-compound-disease network is connected using PageRank and a two-step random walk algorithm. The more components a drug contains, and the greater the impact of these components on the target, the more disease targets it targets, and the more important these targets are in the PPI network, the more effective the drug is in treating the disease, and the higher its H-score. The specific calculation formula is as follows:

[0061]

[0062] In the formula, the overlap rate is the area overlap rate. n represents the number of genes in a drug. Hub-D The number of genes representing a disease. This indicates the number of genes that overlap between the drug and the disease gene sequence.

[0063]

[0064] In the formula, Hscoreγ i To score the effectiveness of a certain drug against a disease, A β V represents the target score for β. βγ This represents the matrix constructed between β and γ, where β represents the degree value of a certain drug and γ represents the degree value of the disease, with 1 and 0 indicating whether there is a connection between them. out This represents the out-degree value of the drug node β in the network.

[0065] Furthermore, based on the network relationship between the disease and the drug target, the network distance between the disease and the drug target is obtained, including:

[0066] Based on the similarity of the network distance of disease genes, the network relationship between diseases and drug targets is obtained; based on the network relationship between diseases and drug targets, the average shortest distance within the group and the average shortest distance between disease-drug target pairs are obtained, wherein drug A, drug B and disease are grouped together.

[0067] The network distance between diseases and drug targets is obtained based on the average shortest distance within a group and the average shortest distance between disease-drug target pairs.

[0068] Specifically, based on the fact that drug-corresponding genes form modules in the protein-protein interactionome, and using the similarity of their network distances to disease genes as background data, the network relationship between disease and drug targets is calculated to obtain the average shortest distance s within the group. AB The average shortest distance d between the disease-drug target pair AB ;

[0069]

[0070] In the formula, d AA and d BB This represents the average shortest distance within the interaction zone between nodes A and B. When s AB When s < 0, the two sets of target points overlap topologically; when s AB When ≥0, the two sets of target points are topologically separated.

[0071] The Z-score is a reliable indicator of 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 using 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-based perspective, if the drug target and the disease target are separated from each other, then the corresponding z≥0; otherwise, z<0.

[0074] Furthermore, gaining experience with medication use includes:

[0075] Based on the database, obtain drug pairs;

[0076] Calculate the Pscore and Mscore values ​​for the corresponding drug pairs, and obtain medication experience based on the Pscore and Mscore values.

[0077] Specifically, by retrieving, deduplicating, and organizing 28,279 drug pairs from the database of traditional Chinese medicine value assessment information, the Pscore value of each drug pair was calculated. All drug compound prescriptions related to clinically diagnosed diseases retrieved within the past 10 years were added. After pre-duplicate removal and normalization, factor analysis was performed, and the Mscore value was used to represent the average frequency of a particular drug occurrence in commonly used drug pairs.

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

[0079]

[0080] In the formula, Herb i and Herb j Both refer to drugs, Pscore j This indicates that a search in the database found a result containing the name of a specific drug, Herb. j The sum of P-scores for all drug pairs, where ε represents the sum of the P-scores for a particular drug Herb after factor analysis. i Maximum absolute value factor weight, n j The value indicates the number of drug pairs retrieved.

[0081] Furthermore, based on the chemical composition of the drug and the disease target, the compound categories obtained include:

[0082] Obtain the chemical components of the drug, perform molecular docking between chemical components that appear at a preset frequency and disease targets, and obtain the molecular fingerprint spectrum after docking.

[0083] Based on molecular fingerprinting, obtain molecular fingerprint similarity;

[0084] The molecular fingerprint similarity is classified to obtain the compound category.

[0085] Specifically, molecular docking was performed on compounds appearing ≥5 times with disease genes. Pymol software was used to dehydrate and remove ligands from the receptor protein, while AutoDockTools software was used to hydrogenate and balance the charge of the receptor protein. AutoDock Vina was used for molecular docking scoring. Results were output as affinity, an important indicator of whether the ligand can effectively bind to the receptor molecule. Molecular fingerprints of effective compounds were obtained using the "rcdk", "ChemmineR", and "cluster" packages. The Tanimoto coefficient was used to measure molecular fingerprint similarity, and K-means clustering was used to classify compounds within the drug into three categories based on their distance.

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

[0087] The input layer is used to input the effective region, network distance, medication experience, and compound category.

[0088] The GCN layer is used to aggregate the feature representations of nodes (or edges) in the graph after preprocessing and update them. In other words, it constructs a matrix to capture the structural information of the graph and the relationship between node features.

[0089] The prediction layer is used to predict the correlation value of drugs.

[0090] The output layer is used to output the correlation values ​​of drugs.

[0091] Specifically, the graph data used consists of disease gene nodes and drug nodes, where each node represents a gene and edges represent interactions between genes. To construct such a graph structure, exemplary graph data is first generated, containing randomly generated node features and edge indices. Furthermore, each edge is assigned a binary label (0 or 1) to indicate its presence. To enhance model interpretability, an inner product operation is performed on each edge using the embeddings of its two endpoints, yielding a scalar as the edge's importance score. The score is then mapped to the [0,1] interval using the Sigmoid function, representing the probability of the edge's existence. 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 zeroed out, and the probability distribution of the edges is calculated based on the current input, with the loss calculated accordingly. The model parameters are updated after backpropagation of the loss. Gradient clipping is incorporated during training to limit the maximum norm to no more than 1.0, as shown in the following formula:

[0092]

[0093] H (l+1) =(AH (l) ·G (l) )

[0094]

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

[0096] Furthermore, based on the correlation value, the drug correlation coefficient is obtained as follows:

[0097] Based on the correlation degree value, the drug correlation coefficient is obtained using the singular value decomposition method.

[0098] Furthermore, using singular value decomposition, the drug correlation coefficients are obtained, including:

[0099] Singular value decomposition is performed on the matrix to obtain the corresponding principal components;

[0100] The correlation coefficients between the corresponding principal components are calculated using the vector inner product to obtain the drug correlation coefficients.

[0101] Specifically, Singular Value Decomposition (SVD) is a common linear transformation method widely used in statistics and machine learning. After information normalization, SVD is applied to the matrix to obtain the corresponding principal components. The correlation coefficients between drugs are then calculated using vector inner products. This method innovatively considers the control capabilities of drugs on disease and compound nodes based on their inherent characteristics. The drugs predicted by the graph convolutional network model are then ranked based on the pairwise correlation coefficients between drugs under SVD, and drug combinations for treating diseases are selected according to degree values ​​and the average correlation coefficient of the drugs.

[0102] The following is in conjunction with the appendix Figure 2 This embodiment will be described in detail:

[0103] This embodiment takes the screening of drugs 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 was identified. Each traditional Chinese medicine (TCM) was scored using overlap rate calculation and H-score, and 109 TCMs were selected based on thresholds of overlap rate > 0 and H-score > 0.001. A PPI network was constructed for the targets of each TCM, and the network proximity calculation from the TCM targets to the disease modules revealed topological intersections (Z < 0) between the TCM targets and the hub genes. Using non-steroidal anti-inflammatory drugs (NSAIDs) as a threshold, 105 TCMs were selected. Based on TCM usage experience, further screening using Mscore > 5 resulted in 45 TCMs, containing a total of 612 compounds.

[0105] A set of effective chemical components for treating diseases was used as the background data set. Combining the overlap between the genes corresponding to traditional Chinese medicine and ankylosing spondylitis, and molecular docking results, a total of 14 effective compounds were identified.

[0106] A similarity compound and drug group space was calculated. Hierarchical and K-means clustering methods were used to further analyze 32 traditional Chinese medicines with ≥3 effective compounds, and 24 of these were included in subsequent graph convolutional network model predictions. Each compound and drug space included background data and traditional Chinese medicine chemical composition data.

[0107] A graph convolutional model was used to predict drug combinations for treating ankylosing spondylitis. The parameters of the GCN model were pre-tuned, and the final running parameters were set as follows: hidden_dim = 32, output_dim = 1, learning_rate = 0.0001, epochs = 100. The training curve showed an initial decrease followed by a plateau, consistent with the expected optimization trend. Small fluctuations in loss values ​​persisted in some epochs, possibly due to data complexity or insufficient gradient updates. A total of 225 predicted traditional Chinese medicine (TCM) results were obtained, of which 146 entries showed potential synergistic relationships between the drugs. By calculating the mean association value of each TCM under SVD, and using association scores > 10 and association coefficients > 0.5, as shown in Table 1, eight TCM drugs were ultimately identified: Myrrh, Drynaria fortunei, Lycium barbarum, Epimedium, Achyranthes bidentata, Alpinia officinarum, Forsythia suspensa, and Astragalus membranaceus.

[0108] Table 1

[0109]

[0110] As shown in Table 2, a 3x cross-validation (CV) strategy was used to compare the model with seven common machine learning models: decision trees, random forests, gradient boosting trees, support vector machines, extreme gradient boosting, lightweight gradient boosting machines, and K-nearest neighbors. The primary outcome metrics used were classification metrics: accuracy, recall, precision, F1 score, area under the ROC curve, and regression metrics: root mean square error and mean absolute error. Predicted values ​​were compared to true values. The data was split with a 7:3 training to test set ratio.

[0111] Table 2

[0112]

[0113]

[0114] Experimental results show that this proposed method demonstrates superior classification performance compared to various methods, including decision trees, random forests, gradient boosting trees, support vector machines, extreme gradient boosting, lightweight gradient boosting machines, and K-nearest neighbors. In particular, it significantly outperforms other models in key evaluation metrics such as accuracy, recall, precision, and F1 score. Although it does not achieve the lowest values ​​for regression tasks such as root mean square error and mean absolute error, its optimization for the unique properties and clinical applications of each drug demonstrates comprehensive advantages. In conclusion, considering drug specificity and clinical usage, this method not only effectively handles classification problems but also provides a more comprehensive solution for regression analysis.

[0115] The above are merely preferred embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A method for drug combination screening based on a multi-level interpretable graph convolutional network, characterized in that, The method comprises the following steps: According to the correlation between drugs and disease targets, the effective area of the drug on the disease is obtained; According to the correlation between drugs and disease targets, the effective area of the drug on the disease is obtained, which comprises: According to the correlation between drugs and disease targets, the overlap rate and H score of the drug on the disease are obtained; According to the overlap rate and H score, the effective area of the drug on the disease is obtained; According to the network relationship between the disease and the drug target, the network distance between the disease and the drug target is obtained; Obtaining drug use experience; Obtaining drug use experience comprises: Based on the database, the drug pair is obtained; Calculate the Pscore value and Mscore value of the corresponding drug pair, and obtain the drug use experience according to the Pscore value and Mscore value; obtain the compound category according to the chemical composition of the drug and the disease target; According to the chemical composition of the drug and the disease target, the compound category is obtained, which comprises: Obtain the chemical composition of the drug, and perform molecular docking on the chemical composition with a frequency of reaching a preset number and the disease target to obtain the molecular fingerprint spectrum after docking; Based on the molecular fingerprint spectrum, the molecular fingerprint similarity is obtained; Classify the molecular fingerprint similarity to obtain the compound category; The effective area, network distance, drug use experience and compound category are input into a graph convolution network model to obtain the correlation value of the drug; The graph convolution network model comprises: an input layer, a GCN layer, a prediction layer and an output layer; The input layer is used for inputting the effective area, network distance, drug use experience and compound category; The GCN layer is used for aggregating and updating the feature representation of the nodes or edges in the graph after preprocessing the data, i.e. constructing a matrix, to capture the relationship between the structure information of the graph and the node features; The prediction layer is used for predicting the correlation value of the drug; The output layer is used for outputting the correlation value of the drug; Based on the correlation value, the drug correlation coefficient is obtained; According to the correlation coefficient, the drug combination is obtained.

2. The method for drug combination screening based on multi-level interpretable graph convolutional network according to claim 1, characterized in that, According to the network relationship between the disease and the drug target, the network distance between the disease and the drug target is obtained, which comprises: According to the similarity of the network distance of the disease gene, the network relationship between the disease and the drug target is obtained; According to the network relationship between the disease and the drug target, the average shortest distance in the group and the average shortest distance between the disease-drug target pair are obtained; Based on the average shortest distance in the group and the average shortest distance between the disease-drug target pair, the network distance between the disease and the drug target is obtained.

3. The method for drug combination screening based on multi-level interpretable graph convolutional network according to claim 1, characterized in that, Based on the correlation value, the drug correlation coefficient is obtained, which comprises: Based on the correlation value, the singular value decomposition method is used to obtain the drug correlation coefficient.

4. The method for drug combination screening based on multi-level interpretable graph convolutional network according to claim 3, characterized in that, Using the singular value decomposition method, the drug correlation coefficient is obtained, which comprises: The corresponding principal components are obtained by singular value decomposition of the matrix; The correlation coefficient between the corresponding principal components is calculated by vector inner product to obtain the drug correlation coefficient.

Images