Hbv inhibitor screening method based on molecular-gene interaction constrained graph convolutional network

Abstract

The application provides a HBV inhibitor screening method based on a molecule-gene interaction constraint graph convolution network. In view of the limitation of a traditional drug discovery method in processing complex biological data, the application is based on a constructed compound library verified by anti-HBV in-vitro activity, a plurality of gene targets associated with the corresponding compound and an interaction network thereof, a graph data processing capacity of a graph convolution network model is used, and a molecule-gene interaction constraint graph convolution network model is constructed. The model combines an interaction matrix of a target protein corresponding to the gene, a gene feature matrix and a compound activity label, and effectively predicts the biological activity category of the compound. The specific steps include data processing, graph data generation, graph convolution network model training, hyperparameter optimization and model evaluation. The model parameter AUC value is 0.97, and the model effect is good. The application provides a new path and idea for virtual screening of anti-HBV drugs, and has potential application value.

Description

Technical Field

[0001] This invention belongs to the fields of chemistry and biology, and relates to methods for predicting compound activity using graph convolutional networks (GCNs) and their applications, particularly to a method for screening HBV inhibitors based on a molecular-gene interaction-constrained graph convolutional network (MGIC-GCN), and its applications in biopharmaceuticals, drug screening, and precision medicine. Background Technology

[0002] Hepatitis B virus infection is a serious global public health problem. Although existing vaccines and antiviral drugs can effectively inhibit viral replication, they are difficult to completely eliminate the virus, and many patients require long-term treatment. Therefore, the discovery of new anti-HBV drug molecules and the study of their mechanisms of action are urgent. Traditional drug discovery methods, such as high-throughput screening and drug design based on known molecular structures, face many challenges, such as high experimental costs, long cycles, and results being affected by various factors. In recent years, graph convolutional networks (GCNs) have become an important tool for processing biological network data due to their powerful graph data processing capabilities. However, existing GCN models still have limitations when processing complex compound molecule-gene interaction data. To address this, this invention proposes a compound activity prediction method based on MGIC-GCN, aiming to improve the model's prediction accuracy and AUC, providing a new approach for the development of anti-HBV drugs. Summary of the Invention

[0003] The purpose of this invention is to address the shortcomings of existing technologies by proposing a method for screening HBV inhibitors based on a convolutional network of compound molecule-gene interaction constraint graphs.

[0004] To achieve the above objectives, the present invention provides the following technical solution: As a first aspect, a method for screening HBV inhibitors based on molecular-gene interaction constraint graph convolutional networks, comprising the following steps:

[0005] S1. Based on existing databases, collect compounds that can act on multiple HBV-related targets. Specifically, screen from compound libraries for compounds and their targets related to HBV (hepatitis B virus) infection in the CHEMBL and Connectivity Map (CMAP) libraries, along with their related gene target indices. Ensure that the selected compounds are correlated with HBV infection. The multiple HBV-related targets specifically refer to more than three.

[0006] S2, Perform in vitro anti-HBV activity tests and cytotoxicity tests on the above compounds, and construct activity tags for the compounds using anti-HBV activity and cytotoxicity as indicators to establish a dataset.

[0007] S3, Data Processing: Test the extraction of association data between compounds and their genes and targets, constructing a compound-gene interaction graph; based on the interaction information between genes and their targets, construct a target interaction matrix, where each element represents the interaction strength between two targets, and set a threshold to extract the edges of the graph. Extract the features of relevant genes and their targets from the gene-target feature matrix as node features for the graph convolutional network model. Construct a target feature matrix, where each row represents the four-dimensional feature vector of the target.

[0008] S4, Graph Data Generation: For each compound in the dataset, a gene-target network graph is constructed, including node features, edge indices, and edge weights. Nodes in the target network graph represent targets associated with the compound, node features represent the four-dimensional features of the targets, and edges represent the interactions between two targets. Edge indices are created for targets with interaction strength greater than a threshold, and the interaction strength between the two targets is used as the edge weight. The edge weights are then converted into edge types. The generated node features, edge indices, edge types, and activity labels are integrated into a graph data object. Specifically, the interaction strength between the two targets is measured using an interaction score.

[0009] S5, Graph Convolutional Network Model Training: A graph convolutional network model is trained to obtain an anti-HBV activity prediction model. Specifically, graph convolutional layers, global pooling layers, and fully connected layers are used for training to classify the activity of compounds. During model training, the cross-entropy loss function is used to optimize the model parameters. The dataset is randomly divided into a training set, a validation set, and a test set. The training set is used to train the prediction model, the validation set is used to verify the model's performance during training, and the test set serves as an independent validation set to ultimately evaluate the model's performance.

[0010] S6, Hyperparameter Optimization: Optuna is used for hyperparameter optimization, including the number of hidden layer neurons and the learning rate. By searching for the optimal combination of the number of hidden layer neurons and the learning rate, the accuracy of the model is improved.

[0011] S7, Model Evaluation: The prediction models trained using various graph convolutional network models were validated and compared. Specifically, the model performance was evaluated using metrics such as accuracy, F1 score, recall, and AUC. The optimal prediction model was selected, and finally, the advantages of the MGIC-GCN model in the compound activity prediction task were verified.

[0012] S8, HBV inhibitor screening: Based on the anti-HBV activity prediction model, input the test compound for prediction, screen test compounds with anti-HBV activity and no cytotoxicity, and obtain HBV inhibitors.

[0013] Furthermore, in step S1, HBV infection-related gene data are collected from databases such as The Comparative Toxicogenomics Database (CTD) and The National Center for Biotechnology Information (NCBI), and are screened from the FDA_HY-L022 Library compound library to select compounds with more than three HBV infection-related targets in the CHEMBL and CMAP libraries and included in the dataset.

[0014] Furthermore, in step S2, the selected compound was systematically evaluated using HepG2.2.15 cells as an experimental model. The in vitro anti-HBV activity test specifically involved determining the inhibitory effect (inhibition rate) of the compound on HBsAg (hepatitis B surface antigen) and HBeAg (hepatitis B e antigen) secretion; the cytotoxicity test specifically involved determining the cell viability of the compound using the in vitro MTT assay.

[0015] (1) Dataset and Preprocessing: Compound-Target Association Data: The compound-target association data file records the relationship between each compound and multiple targets. Let the compound set be C = {c1, c2, ..., cn}, and each compound ci corresponds to a target set G(ci) = {g1, g2, ..., gm}, where gj represents the target associated with compound ci.

[0016] (2) Activity tag: The activity tag of a compound is derived from the analysis and summarization of activity data and reflects its biological activity category.

[0017] The activity tags for compounds, constructed using anti-HBV activity and cytotoxicity as indicators, are as follows: HBsAg or HBeAg secretion inhibition rate ≥50% indicates anti-HBV activity; cell survival rate ≥50% indicates low cytotoxicity. The activity tags for compounds are {0, 1, 2, 3}, where tag 0 indicates HBsAg or HBeAg secretion inhibition <50% (no anti-HBV activity) and cell survival rate <50% (high cytotoxicity); 1 indicates HBsAg or HBeAg secretion inhibition ≥50% (anti-HBV activity) and cell survival rate <50% (high cytotoxicity); 2 indicates HBsAg or HBeAg secretion inhibition rate ≥50% (anti-HBV activity) and cell survival rate ≥50% (low cytotoxicity); and 3 indicates HBsAg or HBeAg secretion inhibition rate <50% (no anti-HBV activity) and cell survival rate ≥50% (low cytotoxicity). Therefore, this problem is defined as a four-category task.

[0018] (3) Gene-target interaction matrix

[0019] We obtained the target interaction matrix using compound-gene-target interaction data (InteractionScore provided by the StringDB database) and community analysis. The interaction matrix M is a p*p matrix, where p represents the total number of genes, and each element Mij represents the interaction strength between genes gi and gj. To generate graph data for each compound, we extracted a submatrix MG(ci) from the set of genes G(ci) associated with that compound. This submatrix defines the target network for compound ci.

[0020] M G(ci) ={M jk |g j ,g k ∈G(ci)}

[0021] In the interaction matrix, a threshold T = 400 is set, meaning that if Mjk > 400, then the target g j and g k If an edge exists between two points, then no edge is established; otherwise, no edge is established.

[0022] (4) Target feature data

[0023] After clustering the dispersed targets, we set four-dimensional features for each target, including indicators of cytotoxicity and anti-HBV activity. The target feature matrix F is a p*4 matrix, where each row represents a four-dimensional feature of a target. For each compound ci, we extract the feature submatrix F of its related targets from the target feature matrix. G(ci) Each row of this matrix corresponds to a four-dimensional feature vector fj of a target.

[0024] Furthermore, in step S4, during the generation of graph data, target network graph data is generated for each compound.

[0025] (1) Nodes and node characteristics

[0026] In the graph data for each compound, nodes represent targets associated with the compound, and node features are the four-dimensional features of the targets. Let compound c... i The feature matrix of the relevant target is F G(ci) Then the matrix Where m is the compound c i The number of associated targets, where 4 represents the four-dimensional features of each target.

[0027] (2) Edge index and edge weight

[0028] Based on the target interaction matrix M G(ci) We extract the edge index and edge weight for each compound graph data. Let the threshold T be denoted as: Then the edge index Ei is defined as:

[0029] Ei={(g j ,g k )|M jk >T}

[0030] The edge weight Ew is M jk The corresponding value. For the R-GCN model, we further transform the edge weights into edge types Et. The transformation of edge weights into edge types uses a threshold partitioning method. Specifically, let the edge type function be f. ω (M jk ):

[0031]

[0032] The edge type assigned by this function is suitable for R-GCN functions, handling the problem of multiple edge types in R-GCN.

[0033] (3) Graphical data representation

[0034] Graph data can be represented as G = (X, Ei, Et, y), where... is the target node feature matrix; Ei is the edge index, representing target pairs whose interaction strength is greater than the threshold or obtained after transformation; Et is the edge type in the R-GCN model, and y∈(0,1,2,3) is the active label of the compound.

[0035] Furthermore, in step S5, the graph convolutional network model training includes the following steps:

[0036] (1) Graph Data Generation: For each compound ci, the relevant target set G(ci) is extracted; then, a submatrix MG(ci) of G(ci) is extracted from the target interaction matrix, and edge indices and edge types are generated based on the interaction matrix; then, the feature submatrix FG(ci) of the relevant targets is extracted from the target feature matrix; finally, the generated node features, edge indices, edge types, and activity labels are integrated into a graph data object. Using the above method, we constructed graph data for GCN, GIN, GAT, and R-GCN models, and used them to train the models to predict the activity categories of compounds.

[0037] (2) Model Definition: Four graph convolutional network models are used: GCN (Graph Convolutional Network), GAT (Graph Attention Network), GIN (Graph Isomorphism Network), and RGCN (Relational Graph Convolutional Network). These four models share a similar overall architecture, including graph convolutional layers, global pooling operations, and fully connected classification layers. Their basic operations can be summarized by a message passing mechanism: the features of a node interact with the features of its neighboring nodes, updating the node representation through convolutional operations.

[0038] The node features in a graph convolutional network are represented as a matrix. Where N is the number of nodes and F is the dimension of node features. The graph structure is represented by an adjacency matrix. Description, where A ij This indicates whether a connection exists between node i and node j. Generally, the basic convolution operation of a graph convolutional network can be abstracted as follows:

[0039] H (l+1) =f(H (l) A)

[0040] Where H (l) This represents the feature matrix of the nodes in the l-th layer, with the initial feature matrix being H. (0) =X, where f is the convolution operation function that aggregates the features of nodes and their neighbors and updates the feature representation through nonlinear transformations. After several layers of graph convolution operations, we perform global pooling operations (such as global average pooling, global max pooling, etc.) on all nodes to generate the overall representation of the graph. Global pooling aggregates the features of all nodes into a single vector h. g :

[0041] h g =GMP(H (L) )

[0042] Where H (L)This is the node representation after the last convolutional layer. Finally, the pooled graph representation is classified through one or more fully connected layers, and the model's output is optimized using the cross-entropy loss function.

[0043]

[0044] Where C represents the number of categories, y i For the true category label, p i The model predicts probabilities. Based on the above architecture, the main difference between the four models lies in the specific implementation of their convolution operations.

[0045] Furthermore, in step S6, hyperparameter optimization, the selection of hyperparameters has a crucial impact on the model's performance. In this study, Optuna was used as a hyperparameter optimization tool, and the optimal model configuration was automatically searched out through multiple trials.

[0046] (1) Definition of search space: Hidden Channels: The number of hidden layer neurons in the model. This value affects the feature extraction capability of each convolutional layer, and the range is 8≤hidden_channels≤32.

[0047] Learning Rate: The learning rate controls the step size for updating model weights. An excessively large learning rate may lead to model instability, while an excessively small learning rate will slow down model convergence. The search range for the learning rate is 10. -6 ≤learning_rate≤10 -3 Its logarithm is defined as: learning_rate = 10log_learning_rate, log_learning_rate ∈ [-6, -3].

[0048] (2) Objective Function: In each trial, Optuna extracts a set of hyperparameters from the search space, applies them to the model, and evaluates the model's performance. This performance metric (usually the accuracy on the test set) serves as the output of the objective function, which is used to adjust the hyperparameters in subsequent iterations. The objective function is defined as: accuracy = f(hidden_channels, learning_rate), where f represents the model training and evaluation process. Hyperparameter Search Process: The hyperparameters are optimized by running multiple trials using the Optuna framework. In each trial, Optuna selects a new set of hyperparameters from the search space and calculates the accuracy on the test set through the training process.

[0049] Furthermore, in step S7, model evaluation, a comprehensive comparison of the performance of the MGIC series graph convolutional models with other mainstream machine learning models was conducted. All models used the same sample data but employed different data processing methods. Evaluation metrics included accuracy, F1 score, recall, and AUC.

[0050] As a second aspect, the present invention provides HBV inhibitors obtained by screening based on molecular-gene interaction constraint graph convolutional networks, including Prostaglandin E2, Febuxostat, and Baricitinib, with the following structural formulas:

[0051]

[0052] The beneficial effects of this invention include: addressing the limitations of traditional drug discovery methods in processing complex biological data, this invention utilizes the graph data processing capabilities of Graph Convolutional Networks (GCNs) to construct the MGIC-GCN model. This model, based on a constructed library of compounds with in vitro anti-HBV activity validation and an index library of compounds and HBV infection-related targets, combines the interaction matrix of target proteins, the feature matrix of target proteins, and the activity tags of compounds to effectively predict the biological activity categories of compounds. Specific steps include data processing, graph data generation, graph convolutional network model training, hyperparameter optimization, and model evaluation. The model parameter AUC value is 0.97, indicating good model performance. This invention provides a new path and approach for virtual screening of anti-HBV drugs and has potential application value. Attached Figure Description

[0053] Figure 1 This is a structural diagram of the MGIC-GCN model. Detailed Implementation

[0054] This invention establishes a dataset of compounds' anti-HBV activity, collecting data on the association between compounds and targets, interactions between targets, and target feature data to construct a target interaction network graph. Subsequently, three graph convolutional network models—GCN, GIN, and RGCN—are selected as training models for the MGIC series of graph convolutional network models to mine compound activity patterns. Experimental verification shows that, compared with traditional machine learning models, MGIC-GCN exhibits significant advantages in accuracy and AUC, and can more accurately handle complex compound molecule-gene interaction graph structures, effectively improving the accuracy of compound activity prediction.

[0055] The following combination Figure 1 The technical solution and beneficial effects of the present invention will be described in detail. Any content not mentioned herein is existing technology or content that can be set or adjusted by those skilled in the art as needed.

[0056] This invention proposes a method for predicting the activity of MGIC-GCN compounds, which specifically includes the following steps:

[0057] S1. 12,200 genes associated with HBV (Hepatitis B virus) infection were collected from databases such as the Comparative Toxicogenomics Database (CTD) and the National Center for Biotechnology Information (NCBI). Compounds were screened from the FDA_HY-L022 Library, selecting compounds with more than three HBV infection-related genes and their targets in the CHEMBL and CMAP libraries. The selected compounds were ensured to be associated with HBV infection. 105 compounds met the criteria.

[0058] S2, In vitro anti-HBV activity assay for these 105 compounds: The sample solution was prepared by dissolving the aforementioned 105 compounds in DMSO at a concentration of 30 μM. The effects of the compounds on HBeAg and HBsAg antigen secretion were determined using ELISA. The specific experimental procedures are as follows: First, wash HepG2.2.15 cells in suitable condition with PBS, add trypsin to disperse the cells, add culture medium (MEM + 10% FBS + 380 mg / mL G418) and pipette to form a single-cell suspension; then, transfer the suspension to an EP tube (15 mL), centrifuge to remove the supernatant, add 2 mL of complete culture medium (MEM + 10% FBS + 380 μg / mL G418) and pipette to form a single-cell suspension, re-seed into 48-well plates, with approximately 3000 cells per well, and incubate at 37℃, 5% CO2 for 24 hours; finally, aspirate the culture medium and treat the cells with the prepared sample solution (final concentration 30 μM), using lamivudine as a positive control and cells without sample treatment as a blank control, incubate at 37℃, 5% CO2 for 72 hours and collect the culture medium. The HBeAg detection kit (Shanghai Kehua Bioengineering Co., Ltd., National Medical Device Registration Certificate No. 20163400144) and the HBsAg detection kit (Shanghai Kehua Bioengineering Co., Ltd., National Drug Approval No. S10910113) were used to detect the samples. The absorbance (OD value) was measured using an ELISA reader at a detection wavelength of 450 nm and a reference wavelength of 630 nm. The inhibition rates of the compounds against HBsAg and HBeAg were calculated according to the following formula:

[0059]

[0060] S3, In vitro MTT assay to determine the cell viability of the compound.

[0061] (1) Wash HepG2.2.15 cells in suitable cell condition with PBS, remove cell culture medium, add 0.25% trypsin to digest cells, add culture medium (MEM+10%FBS+380mg / mL G418) and pipette to form a single cell suspension; then, transfer the suspension to EP tube (15mL), centrifuge to remove supernatant, add 2mL of complete culture medium (MEM+10%FBS+380mg / mL G418) and pipette to form a single cell suspension, and re-seed in 96-well plates, with about 3000 cells per well;

[0062] (2) After 12 hours of adhesion, remove the culture medium and add 100 μL of culture medium (MEM+10%EBS+380mg / mLG418) containing different concentrations of the test sample (serially diluted from below 200μM to 1.56μM). Also set up blank wells (containing only culture medium) and control wells (without adding solvent of the test sample).

[0063] (3) After the cells and drugs were co-incubated at 37°C and 5% CO2 for 24 hours, 10 μL of MTT solution (10 mg / mL) was added to each well and co-incubated for 4 hours. Then, 100 μL of MTT solution was added and the cells were incubated at 37°C overnight. The absorbance (A value) of each well was measured at 550 nm.

[0064] (4) Calculate cell viability. Viability = (experimental wells - blank wells) / (control wells - blank wells) × 100%. The experiment was repeated three times.

[0065] Anti-HBV activity analysis (Table 1) showed that 38 small molecules inhibited HBsAg secretion by cells by more than 50%, and 2 small molecules inhibited HBeAg secretion by more than 50%. Among them, 11 compounds showed cell viability greater than 50% as determined by MTT assay, indicating that their anti-HBV activity did not affect cell viability.

[0066] Table 1. Results of HBV inhibitor activity

[0067]

[0068]

[0069] Table 2. Molecules of Anti-HBV Active Compounds

[0070]

[0071]

[0072]

[0073]

[0074] S4, Dataset and Preprocessing: To construct a Graph Convolutional Network (GCN) model for predicting compound activity, we first extracted information on the association between compounds and targets, interaction data between targets, and target feature data from multi-source data. Using this data, we generated a target interaction network graph for each compound. In each network graph, nodes represent targets, edges represent interactions between targets, and node features are the four-dimensional features of the targets. The ultimate goal is to predict the activity category of compounds using graph data.

[0075] (1) Based on the CHEMBL and CMAP libraries, the association data between compounds (Table 3) and targets: The association data file between compounds and targets records the relationship between each compound and multiple targets. Let the set of compounds be C = {c1, c2, ..., cn}, and each compound ci corresponds to a target set G(ci) = {g1, g2, ..., gm}, where gj represents the target associated with compound ci.

[0076] (2) Activity Tags: The activity tags of compounds are derived from the analysis and summarization of activity data, reflecting their biological activity categories. An inhibition rate of ≥50% for HBsAg and HBeAg secretion indicates inhibitory activity; a cell survival rate of ≥50% indicates low cytotoxicity. Specifically, tag 0 indicates HBsAg and HBeAg secretion inhibition <50% and cell survival rate <50%; 1 indicates HBsAg and HBeAg secretion inhibition ≥50% and cell survival rate <50%; 2 indicates HBsAg and HBeAg secretion inhibition ≥50% and cell survival rate ≥50%; and 3 indicates HBsAg and HBeAg secretion inhibition <50% and cell survival rate ≥50%. Therefore, this problem is defined as a four-category task.

[0077] (3) Target interaction matrix

[0078] We obtained the target interaction matrix through compound target community analysis. The interaction matrix M is a p*p matrix, where p represents the number of all targets, and each element Mij in the matrix represents the interaction strength between targets gi and gj. This invention uses the interaction score provided by the stringDB database as a measure of the interaction strength between targets. To generate graph data for each compound, we extract a submatrix MG(ci) of the target set G(ci) associated with that compound. This submatrix defines the target network of compound ci.

[0079] M G(ci) ={M jk |g j ,g k ∈G(ci)}

[0080] In the interaction matrix, a threshold T = 400 is set, meaning that if Mjk > 400, then the target g j and g k If an edge exists between two points, then no edge is established; otherwise, no edge is established.

[0081] (4) Target feature data

[0082] After clustering the dispersed targets, we set a four-dimensional feature for each target, including toxicity and activity indicators. This invention uses four labels (0, 1, 2, 3) for anti-HBV activity data. The target feature matrix F is a p*4 matrix, where each row represents a four-dimensional feature of a target. For each compound ci, we extract the feature submatrix F of its related targets from the target feature matrix. G(ci) Each row of this matrix corresponds to a four-dimensional feature vector fj of a target.

[0083] S5, In the generation of graph data, based on the above steps, target network graph data is generated for each compound.

[0084] (1) Nodes and node characteristics

[0085] In the graph data for each compound, nodes represent targets associated with the compound, and node features are the four-dimensional features of the targets. Let compound c... i The feature matrix of the relevant target is F G(ci) Then the matrix Where m is the compound c i The number of associated targets, where 4 represents the four-dimensional features of each target.

[0086] (2) Edge index and edge weight

[0087] Based on the target interaction matrix M G(ci) We extract the edge index and edge weight for each compound graph data. Let the threshold T be denoted as: Then the edge index Ei is defined as:

[0088] Ei={(g j ,g k )|M jk >T}

[0089] The edge weight Ew is M jk The corresponding value. For the R-GCN model, we further transform the edge weights into edge types Et, and let the edge type function be f. ω (M jk ):

[0090]

[0091] The edge type assigned by this function is suitable for R-GCN functions, handling the problem of multiple edge types in R-GCN.

[0092] (3) Graphical data representation

[0093] Graph data can be represented as G = (X, Ei, Et, y), where... is the target node feature matrix; Ei is the edge index, representing target pairs whose interaction strength is greater than the threshold or obtained after transformation; Et is the edge type in the R-GCN model, and y∈(0,1,2,3) is the active label of the compound.

[0094] S6, Training a graph convolutional network model includes the following steps:

[0095] (1) Graph Data Generation: First, for each compound ci, the relevant target set G(ci) is extracted; then, a submatrix MG(ci) of G(ci) is extracted from the target interaction matrix, and edge indices and edge types are generated based on the interaction matrix; then, the feature submatrix FG(ci) of the relevant targets is extracted from the target feature matrix; finally, the generated node features, edge indices, edge types, and activity labels are integrated into a graph data object. Using the above method, we constructed graph data for GCN, GIN, GAT, and R-GCN models, and used them to train the models to predict the activity categories of compounds.

[0096] (2) Definition of MGIC series graph convolutional network models: Four graph convolutional network models are used, namely GCN (Graph Convolutional Network), GAT (Graph Attention Network), GIN (Graph Isomorphism Network), and RGCN (Relational Graph Convolutional Network). Although these models have their own innovative features, they share a similar structure in their overall architecture, including graph convolutional layers, global pooling operations, and fully connected classification layers. Their basic operations can be summarized by a message passing mechanism, that is, the features of a node interact with the features of its neighboring nodes, and the node representation is updated through convolutional operations.

[0097] The node features in a graph convolutional network are represented as a matrix. Where N is the number of nodes and F is the dimension of node features. The graph structure is represented by an adjacency matrix. Description, where A ij This indicates whether a connection exists between node i and node j. Generally, the basic convolution operation of a graph convolutional network can be abstracted as follows:

[0098] H (l+1) =f(H (l)A)

[0099] Where H (l) This represents the feature matrix of the nodes in the l-th layer, with the initial feature matrix being H. (0) =X, f is the convolution operation function, which aggregates the features of nodes and their neighbors and updates the feature representation through nonlinear transformation.

[0100] After several layers of graph convolution operations, we perform global pooling operations (such as global average pooling, global max pooling, etc.) on all nodes to generate a global representation of the graph. Global pooling aggregates the features of all nodes into a single vector h. g :

[0101] h g =GMP(H (L) )

[0102] Where H (L) This is the node representation after the last convolutional layer. Finally, the pooled graph representation is classified through one or more fully connected layers, and the model's output is optimized using the cross-entropy loss function.

[0103]

[0104] Where C represents the number of categories, y i For the true category label, p i The model predicts probabilities. Based on the architecture described above, the main difference between the four models lies in the specific implementation of their convolution operations. The unique features of each model are described below.

[0105] Graph Convolutional Networks (GCNs): These networks utilize convolution operations based on spectral graph theory. Their convolution operations are implemented by normalizing the adjacency matrix, which is defined as follows:

[0106]

[0107] in This involves adding self-loops (connections between a node and itself) to the original adjacency matrix, where D is the corresponding degree matrix. The convolution operation in GCN is represented as:

[0108] H (l+1) =σ(AH (l) W (l) )

[0109] Among them W (l) Let σ be the learnable weight matrix of the l-th layer, and σ be the activation function (usually ReLU). GCN uses a normalized adjacency matrix to ensure that feature aggregation considers the features of the node itself and its neighbors, while preventing the gradient vanishing problem caused by excessively large degree values.

[0110] Graph Attention Networks (GAT) introduce a self-attention mechanism, allowing the model to assign different weights to different neighbor nodes when aggregating features. Unlike GCNs, which use a fixed adjacency matrix, GAT's attention mechanism enables each node to dynamically decide which neighbors to focus on. Specifically, GAT uses learnable attention coefficients α. ij This represents the correlation between node i and node j. This coefficient is calculated using the following formula:

[0111]

[0112] in:

[0113]

[0114] z ij =a T [Wh i ||Wh j ]

[0115] Where a is the weight vector used to calculate the attention score, W is the feature transformation matrix, || denotes the concatenation operation, and h i and h j Let N(i) be the feature vectors of node i and node j, LeakyReLU be an activation function, and N(i) be the set of neighboring nodes of node i. The final node feature aggregation is accomplished through a multi-head attention mechanism.

[0116]

[0117] Where K is the number of attention heads, and ∪ represents the splicing operation.

[0118] Graph Isomorphic Networks (GINs): Designed to have a stronger ability to distinguish graph structures, the goal is to enable the model to have unique representation capabilities when dealing with different graph structures. The convolution operation of GINs is similar to weighted summation, updating node representations by aggregating features of the node itself and its neighbors. The specific formula for GIN convolution is:

[0119]

[0120] Here, ∈ is a learnable or fixed scalar parameter, and MLP (Multilayer Perceptron) is used to perform nonlinear transformations on the aggregated features. The innovation of GIN lies in using MLP to enhance feature representation capabilities and in using ∈ adjustment to enable the model to flexibly handle combinations of node features and neighbor features.

[0121] Relational Graph Convolutional Networks (RGCNs) are used to handle graph structures with heterogeneous relationships, where edges between nodes can represent different types of relationships. To this end, RGCNs design independent convolutional kernel weights Wr for each relationship type r. The convolution operation of RGCNs is represented as follows:

[0122]

[0123] S7. In hyperparameter optimization, the selection of hyperparameters has a crucial impact on model performance. In this study, Optuna was used as a hyperparameter optimization tool, and the optimal model configuration was automatically searched through multiple trials.

[0124] (1) Definition of search space:

[0125] Hidden Channels: The number of neurons in the hidden layers of the model. This value affects the feature extraction capability of each convolutional layer, and its range is 8 ≤ hidden_channels ≤ 32.

[0126] Learning Rate: The learning rate controls the step size for updating model weights. An excessively large learning rate may lead to model instability, while an excessively small learning rate will slow down model convergence. The search range for the learning rate is 10. -6 ≤learning_rate≤10 -3 Its logarithm is defined as: learning_rate = 10log_learning_rate, log_learning_rate ∈ [-6, -3].

[0127] (2) Objective function:

[0128] In each trial, Optuna extracts a set of hyperparameters from the search space, applies them to the model, and evaluates the model's performance. This performance metric (typically the accuracy on the test set) is output as the objective function, used to adjust the hyperparameters in subsequent iterations. The objective function is defined as: accuracy = f(hidden_channels, learning_rate), where f represents the model training and evaluation process.

[0129] Hyperparameter search process: Using the Optuna framework, we ran multiple trials to optimize the hyperparameters. In each trial, Optuna selected a new set of hyperparameters from the search space and calculated the accuracy on the test set through the training process.

[0130] Ultimately, we obtained the optimal combination of hyperparameters. For example, the optimized number of hidden channels is 19, and the learning rate is 7.244 × 10⁻⁶. -5 .

[0131] S8, Model Training and Evaluation: A comprehensive performance comparison was conducted between the four MGIC graph convolutional network models and other mainstream machine learning models. All models used the same sample data but employed different data processing methods. Evaluation metrics included accuracy, F1 score, recall, and AUC. Based on the experimental results, the GCIC series demonstrated significant advantages, especially in accuracy and AUC. Specifically, the GCIC series models achieved an accuracy of 0.9137 and an AUC of 0.9787, indicating that GCIC has very strong predictive ability in this task, effectively classifying samples and distinguishing between positive and negative samples well. Compared with traditional machine learning models, GCIC significantly outperformed models such as XGBoost, Support Vector Machine (SVM), and Random Forest (RF) in both accuracy and AUC. For example, XGBoost had an accuracy of 0.7692 and an AUC of 0.5054; SVM had an accuracy of 0.7692 and an AUC of 0.5830; while Random Forest performed the weakest, with an accuracy of 0.7308 and an AUC of 0.6182. This indicates that the GCIC series is better able to handle complex data structures, especially graph-structured data, demonstrating its strong advantage in this task. In comparison, the performance gap between GCIC and DeepBindGCN is relatively small. GCIC's accuracy is 0.9021 and its AUC is 0.9648, performing close to DeepBindGCN. This shows that the GeneNet series is competitive with DeepBindGCN in terms of accuracy and task performance stability.

[0132] Table 3. Model Parameter Table

[0133]

[0134] S9 Model Prediction and Activity Validation: The MGIC-GCN model was used to predict the activity of compounds not appearing in the molecular-gene network dataset to determine whether a compound met the tag "2" condition, where "2" indicates an inhibition rate of HBsAg secretion and HBeAg secretion ≥50% and a cell survival rate ≥50%. Thirteen compounds were tested, and the validation results are shown in Table 4. Five compounds showed anti-HBV activity, of which three inhibited HBeAg secretion, and three of these compounds showed a cell survival rate ≥50%. The relevant structures are shown in Table 5.

[0135] Table 4. Prediction and Validation Results of the MGIC-GCN Model

[0136]

[0137] Table 5. Compound structures predicted and verified by the MGIC-GCN model

[0138]

[0139] In summary, this invention proposes a method for screening HBV inhibitors based on a molecular-gene interaction constraint graph convolutional network. The model parameter AUC value is 0.97, demonstrating good performance, and novel HBV inhibitors were successfully screened. This invention provides a new pathway and approach for virtual screening of anti-HBV drugs and has potential application value.

Claims

1. A method for screening HBV inhibitors based on molecular-gene interaction constraint graph convolutional networks, characterized in that, Includes the following steps: Based on existing databases, compounds that can act on multiple HBV-related targets were collected, and in vitro anti-HBV activity and cytotoxicity tests were performed. The activity tags of the compounds were constructed using anti-HBV activity and cytotoxicity as indicators, and a dataset was established. Construct a target interaction matrix, where each element represents the interaction strength between two targets; construct a target feature matrix, where each row represents a four-dimensional feature vector of the target. For each compound in the dataset, nodes are set as targets associated with the compound, and node features are four-dimensional features of the target, including toxicity and activity indicators. HBV activity tags are used to construct the four-dimensional features of the targets. For targets where the interaction strength between two targets is greater than a threshold, an edge index is established, and the interaction strength between the two targets is used as the edge weight. The edge weight is then converted into an edge type. The generated node features, edge indexes, edge types, and activity tags are integrated into a graph data object. A graph convolutional network model is trained to obtain an anti-HBV activity prediction model. The test compound is input for prediction, and test compounds with anti-HBV activity and no cytotoxicity are screened to obtain HBV inhibitors. The multiple HBV-related targets specifically refer to more than three. The in vitro anti-HBV activity test specifically involves determining the inhibition rate of the compound against HBsAg and HBeAg. The cytotoxicity test specifically involves determining the cell viability of the compound using the in vitro MTT assay. The activity tags for compounds, constructed using anti-HBV activity and cytotoxicity as indicators, are as follows: HBsAg or HBeAg inhibition rate greater than or equal to 50% indicates anti-HBV activity, and cell survival rate greater than or equal to 50% indicates low cytotoxicity. The activity tags for compounds are {0, 1, 2, 3}, where 0 indicates no anti-HBV activity and high cytotoxicity, 1 indicates anti-HBV activity and high cytotoxicity, 2 indicates anti-HBV activity and low cytotoxicity, and 3 indicates no anti-HBV activity and low cytotoxicity.

2. The HBV inhibitor screening method based on molecular-gene interaction constraint graph convolutional network as described in claim 1, characterized in that, The conversion of edge weights into edge types is achieved using a threshold partitioning method.

3. The HBV inhibitor screening method based on molecular-gene interaction constraint graph convolutional networks as described in claim 1, characterized in that, The method for training an anti-HBV activity prediction model based on a graph convolutional network model involves the following steps: the graph convolutional network model is trained using graph convolutional layers, global pooling layers, and fully connected layers, and the model parameters are optimized using a cross-entropy loss function; the dataset is randomly divided into a training set, a validation set, and a test set. The training set is used to train the prediction model, the validation set is used to verify the model's performance during training, and the test set serves as an independent validation set to ultimately evaluate the model's performance.

4. The HBV inhibitor screening method based on molecular-gene interaction constraint graph convolutional network as described in claim 1, characterized in that, The training of the graph convolutional network model to obtain the anti-HBV activity prediction model also includes: using Optuna for hyperparameter optimization, whereby the optimized hyperparameters include the number of hidden layer neurons and the learning rate.

5. The HBV inhibitor screening method based on molecular-gene interaction constraint graph convolutional network as described in claim 1, characterized in that, The method of training a graph convolutional network model to obtain an anti-HBV activity prediction model also includes: verifying and comparing prediction models trained using multiple graph convolutional network models, and selecting the optimal prediction model based on metrics such as Accuracy, F1Score, Recall, and AUC.

6. The HBV inhibitor screening method based on molecular-gene interaction constraint graph convolutional network as described in claim 1, characterized in that, The interaction strength between the two targets is measured by the interaction score.

7. An HBV inhibitor obtained using the HBV inhibitor screening method based on molecular-gene interaction constraint graph convolutional networks as described in any one of claims 1-6, characterized in that, Including Prostaglandin E2, Febuxostat, and Baricitinib, the structural formulas are as follows: 。

Images