Drug-target interaction prediction method based on hierarchical representation learning framework

Abstract

The application provides a drug-target correlation prediction method based on a hierarchical representation learning framework, in the method, high information density nodes are screened based on a dynamic fluctuation threshold, low noise nodes are reduced, and topology connection is reconstructed according to a virtual edge weight formula; drug molecule features extracted by a dynamic neighborhood search framework and protein semantic features generated by a denoising autoencoder are spliced in a tensor level to form joint feature representation of a drug-protein pair; a multi-head attention mechanism is used to assign dynamic weights to multi-view features based on the drug-protein pair, multi-view feature vectors are spliced, key information is screened through the attention mechanism, and then the full connection neural network is input, and a correlation prediction value is output based on the full connection neural network. The correlation prediction method solves the problems that the prediction accuracy of the model for complex biological interactions is poor, and the understanding ability of the model for multi-level feature learning correlation in a complex biological network is severely limited.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of protein function prediction, in particular to a drug-target association prediction method based on a hierarchical representation learning framework. BACKGROUND

[0002] Drug development is a complex and time-consuming process, in which identifying drug-target interactions (DTI) is a key step.

[0003] In the field of biological similarity, drug-target interaction (DTI) prediction has become a core topic in computational biochemistry.

[0004] Existing computational methods are mainly divided into three categories: structure-based, ligand-based and machine learning-based:

[0005] The structure-based methodology system is based on the three-dimensional structural features of drug molecules and protein targets, and realizes interaction prediction through molecular docking, molecular dynamics simulation and other technical means. However, the high cost and low throughput characteristics of protein three-dimensional structure analysis technology make this method encounter technical bottlenecks when facing a large number of unanalyzed protein targets, and its application range is severely limited.

[0006] Ligand-based methods, such as this method, analyze the chemical structure characteristics of known active molecules to construct pharmacophore models or quantitative structure-activity relationship (QSAR) models, but in the absence of sufficient known active molecule data support, the prediction performance of such methods will be significantly degraded, and there is a data sparsity bottleneck.

[0007] Machine learning-based methods, such as graph neural networks (GNNs), have shown significant advantages in biological network analysis due to their powerful modeling capabilities for data, and deep learning has achieved remarkable results in bioinformatics.

[0008] However, in the current method for machine learning of graph neural networks (GNNs), the prior art lacks a dynamically adaptive node screening mechanism, making it difficult to effectively distinguish high-value key nodes from low-value noise nodes in networks with sparse node connections and significant information density differences, resulting in the continuous interference of redundant information of low-value nodes in the neighborhood aggregation process, causing feature blurring and over-smoothing, ultimately limiting the prediction accuracy of the model for complex biological interactions. In addition, traditional GCN uses a non-discriminatory average aggregation strategy. When the number of layers increases, the features of low-value noise nodes (such as redundant proteins with non-specific binding) and high-value key nodes are treated equally, causing the features of different functional nodes to gradually converge due to the sharing of multiple layers of low-value neighbors, resulting in a serious over-smoothing problem. Moreover, existing methods cannot effectively model the long-range semantic association between drugs and targets when integrating multi-hop information, resulting in the dilution of long-distance high-value signals by nearby noise, and severely limiting the model's ability to understand multi-level feature learning associations in complex biological networks. SUMMARY

[0009] The purpose of the embodiments of the present application is to provide a drug-target association prediction method based on a hierarchical representation learning framework, to at least solve the technical problems of poor prediction accuracy of the model for complex biological interactions and severely limited understanding ability of the model for multi-level feature learning associations in complex biological networks in the related art.

[0010] According to one embodiment of the present application, a drug-target association prediction method based on a hierarchical representation learning framework is provided.

[0011] The association prediction method comprises the following steps:

[0012] S101. In the constructed heterogeneous graph containing drugs, proteins, diseases and side effects, drug subgraphs and protein subgraphs are extracted according to different meta-paths, and a hierarchical feature learning framework MHGIN based on dynamic neighborhood search is applied to the drug subgraphs and protein subgraphs respectively to learn drug and protein features;

[0013] S102. In each layer of the hierarchical feature learning framework, high information density nodes are screened based on a fluctuation threshold, and TOP-K neighbors are screened, and only high-density nodes filtered by the threshold are retained for calculation. After reducing low-value noise nodes, the topology connection is reconstructed using virtual edges, and all drug-protein pairs DPPs are traversed.

[0014] S103. The drug molecular features extracted by the dynamic neighborhood search framework MHGIN are concatenated with the protein semantic features generated by the denoising autoencoder DAE in a tensor level, the dimension is compressed through a learnable weight matrix, and a joint feature representation of the drug-protein pairs DDPs is formed.

[0015] S104. A multimodal feature fusion strategy is adopted, and a multi-head attention mechanism is used to assign dynamic weights to multi-view features based on drug-protein pairs. The multi-view feature vectors are spliced, and after the key information is filtered by the attention mechanism, the information is input into a fully connected neural network, and the association prediction value is output based on the fully connected neural network.

[0016] Preferably, the hierarchical feature learning framework MHGIN of step S101 includes:

[0017] The underlying layer builds a dynamic neighborhood search module based on feature density and functional correlation, assigning neighborhood weights through a learnable density-aware function to suppress low-degree and low-density noise nodes and strengthen the aggregation of high-density key nodes.

[0018] A cross-layer attention mechanism is introduced in the middle layer to construct a multi-hop path semantic association matrix, and long-range dependencies are explicitly modeled through graph diffusion or path weight optimization.

[0019] The top-level hierarchical feature fuser is designed to adaptively fuse neighborhood features with different hop counts and screening weights according to the semantic level, forming a multi-level feature representation that includes local characteristics, medium-range correlations, and long-range dependencies.

[0020] Preferably, in step S101, for setting the order =1,2,... ; Obtain abstract features through the NGNN layer , expressed as:

[0021]

[0022] in, Indicates the The calculation function of the NGNN layer, , the input feature matrix is , N is the number of samples, d in is the dimension of the input features, represents the set of real numbers, the adjacency matrix ;

[0023] For each abstract feature matrix Find the mean along the sample dimension and get , expressed as:

[0024]

[0025] In the formula, mean represents the mean operation, dim represents the dimension, , and then concatenate these mean vectors in the 0th dimension to get , the dimension of the output feature is ;

[0026] In the attention weight calculation, first pass through the first fully connected layer Transform , get Z1, denoted as:

[0027]

[0028] In the formula, The weight matrix of , T represents the transpose operation of the matrix, the bias vector is , r represents the dimension reduction factor; ReLU(x) = max(0, x) is the rectified linear unit activation function, ;

[0029] Then pass through the second fully connected layer Transform Z1, The weight matrix of , the bias vector is , get: , Finally, use the Softmax function to normalize Z2 along the 0th dimension to get the attention weight matrix, denoted as:

[0030]

[0031] Where, i represents the index of the i-th view in the attention weighting, ; j represents the dimension index of the output feature, j = 1, 2,... d out , , d out represents the dimension of the output feature, represents the number of layers in the graph neural network, and for each column j , assign a probability distribution over different features or feature combinations as the attention weight for each feature dimension indexed by i; k represents the index of the view feature, used for the normalization term of the Softmax denominator, the range is consistent with i; exp represents the calculation of the exponential; z ij represents the value of the i-th row and the j-th column, z kj represents the value of the k-th row and the j-th column.

[0032] In feature weighting and fusion, multiply each abstract feature matrix with the corresponding attention weight vector (weight vector corresponding to the k-th view) element by element to get the weighted feature matrix , denoted as: , where ; finally, add all the weighted feature matrices to get the final output Z, denoted as: wherein k represents a view index, k = 1, 2,..., n view ; Z comprises a drug feature and a protein feature .

[0033] Preferably, in step S102, the node degree is counted by an adjacency matrix A of the drug-protein network in the heterogeneous graph network , for reflecting the number of node connections, the node density is obtained by the number of connections, the degree of each node in each layer network is calculated, the threshold value is set according to the degree distribution characteristics, and the dynamic threshold value is determined according to the inter-class variance.

[0034] Preferably, the step of setting the threshold value according to the degree distribution characteristics and determining the dynamic threshold value according to the inter-class variance comprises:

[0035] The node degree value is counted with L degrees, and the total node degree is N deg , wherein the maximum node degree is D, and the node degree change range is [0, D];

[0036] The number of nodes in each node degree level The probability of each node degree level is calculated , which is represented as: ;

[0037] The threshold value K is preliminarily determined as a percentile , which is represented as , wherein represents the upward rounding function;

[0038] According to the threshold value K, the nodes are divided into two categories:

[0039] Low-degree nodes : ;

[0040] High-degree nodes : ;

[0041] The cumulative probability and mean value of the two intervals are calculated:

[0042] The cumulative probability of the low-degree node class is represented as: ;

[0043] The cumulative probability of the high-degree node class is represented as: ;

[0044] The cumulative mean value of the low-degree node class is represented as: ;

[0045] High node class cumulative mean is expressed as: ;

[0046] Drug-protein node degree total mean is expressed as: ;

[0047] Calculate the inter-class variance is expressed as: ;

[0048] Find the fluctuation threshold So that the difference between nodes with high contact degree and low contact degree is as large as possible, expressed as:

[0049]

[0050] In the formula, Indicates the probability of each node degree level, ε represents each node degree level, K represents the preliminary threshold, and D represents the maximum node degree; Indicates the inter-class variance, Indicates the node degree;

[0051] Fluctuation threshold According to the changes of network structure, the density of node information and the characteristics of data set, the fluctuation threshold is adjusted when the structure of the network changes dramatically, so as to better eliminate redundant nodes and retain key nodes and long-range dependent paths.

[0052] Preferably, in step 102, the step of reconstructing the topology connection by using virtual edges includes: after the low-degree noise nodes are eliminated, connecting them with nodes with high information density in the heterogeneous graph network, and retaining important topological structure information by increasing virtual edges , expressed as: ; In the formula, Indicates the connection weight of the virtual edge, which is assumed to be a constant value and does not perform information transmission, but only transmits topological information.

[0053] Preferably, in step S104, the multi-view includes a topology graph, a semantic graph and a collaboration graph, and the collaboration graph is obtained by averaging the feature representation of the topology graph and the semantic graph, wherein the feature representation calculation formula of the collaboration graph is:

[0054]

[0055] In the formula, Indicates the feature representation of the collaboration graph in the i-th layer; Indicates the topology graph In the i-th layer, the feature representation is denoted as The semantic graph is denoted as In the i-th layer, the feature representation is denoted as

[0056] Preferably, in step 104, a multi-head attention mechanism is used to assign dynamic weights to the drug-protein pair-based multi-view features, and the step of splicing the multi-view feature vectors includes:

[0057] The attention vector is calculated and denoted as:

[0058]

[0059] In the formula, The attention vector is denoted as respectively correspond to the attention scores of the topological feature representation , the semantic feature representation , and the collaborative feature representation ; The attention function is denoted as

[0060] For the joint feature representation of each drug-protein pair in the topological graph, semantic graph, and collaborative graph , the query matrix , key matrix , and value matrix are generated through a learnable linear mapping, where , , is the weight matrix, denotes the real set, and N denotes the number of samples, denotes the dimension of the joint feature representation, and d k denotes the attention dimension;

[0061] Under the multi-head attention mechanism, the attention scores are calculated and weightedly fused by combining multiple Q, , , denoted as:

[0062]

[0063] where is a scaling factor used to normalize the dot product result of the query Q and the key vector;

[0064] The weight score of the topological feature embedding is denoted as:

[0065]

[0066] In the formula, W represents a learnable parameter matrix, used for linear transformation of the topological feature embedding, adjusting the dimension and weight of the feature; b represents a learnable bias vector; q represents a learnable attention vector; represents a hyperbolic tangent activation function, and the result is subjected to nonlinear transformation, and then dot product operation is performed with to obtain the weight score ; the parameter T represents matrix transposition operation;

[0067] Similarly, the weight score of the semantic feature embedding and the weight score of the collaboration feature embedding are calculated;

[0068] The weight scores are normalized to calculate the attention coefficients, denoted as:

[0069]

[0070] In the formula, represents the attention coefficient of the topological feature, represents the weight score of the topological feature, represents the weight score of the semantic feature, represents the weight score of the collaboration feature; The function normalizes , , together, so that the sum of all attention coefficients is 1, and each attention coefficient is in the interval [0, 1];

[0071] Similarly, the attention coefficient of the semantic feature representation and the attention coefficient of the collaboration feature representation are calculated;

[0072] The topological feature representation , , , the semantic feature representation , and the collaboration feature representation are weighted and fused using the attention coefficients calculated in the foregoing to obtain the fused feature representation of the node i.

[0073] Preferably, the step of outputting the correlation prediction value based on the fully connected neural network in step S104 includes:

[0074] The obtained fused feature representation The input is input into the multi-layer fully connected network, and the fully connected neural network learns the complex relationship between the features and the drug-target association by adjusting the weights and biases in the network, and models the prediction task as a binary classification problem:

[0075] When the association score output by the network is greater than the threshold value, it is determined that there is an interaction between the drug and the target;

[0076] When the association score output by the network is not greater than the threshold value, it is determined that there is no interaction.

[0077] Compared with the prior art, the beneficial effects of the drug-target association prediction method based on the hierarchical representation learning framework of the present application are:

[0078] First, the present application constructs a hierarchical feature learning framework that integrates high-density information neighborhood search and low-density node reduction, uses dynamic threshold to accurately filter high information density nodes, and provides high-quality input for multi-dimensional feature space projection matrix; This mechanism retains the feature difference of key functional nodes, effectively avoids the interference of noise nodes on attention allocation, enables the model to dynamically adjust the contribution of each level of information according to node attributes and topological structure, avoids the feature convergence problem, breaks through the bottleneck limitation of "number of layers-performance" in traditional methods, improves the depth of the model while maintaining performance optimization, and provides a new solution for efficient feature learning in complex network scenarios;

[0079] Second, the present application constructs a hierarchical framework MHGIN with dynamic neighborhood search, which realizes efficient integration and intelligent filtering of multi-scale neighborhood information in complex biological networks; Specifically, the framework adopts a three-layer architecture of dynamic neighborhood filtering, cross-layer semantic association and hierarchical fusion to break through the limitations of traditional single-hop neighborhood aggregation through multi-level feature matrix fusion; In addition, the framework introduces an adaptive attention weighting aggregation module to realize differentiated fusion of information of different hop counts by constructing an attention weight matrix, which is generated based on the joint learning of node attribute features and structure features by a multilayer perceptron, effectively solving the feature oversmoothing problem caused by traditional average aggregation. BRIEF DESCRIPTION OF DRAWINGS

[0080] The drawings described herein are used to provide further understanding of the present application, and form a part of the present application. The illustrative embodiments of the present application and their descriptions are used to explain the present application, and do not constitute an improper limitation on the present application.

[0081] In the drawings:

[0082] Figure 1 is the implementation flowchart of the drug-target association prediction method based on the hierarchical representation learning framework according to the embodiment of the present application;

[0083] Figure 2is a subflowchart of a drug-target association prediction method based on a hierarchical representation learning framework according to an embodiment of the present application;

[0084] Figure 3 is a structural diagram of a drug-target association prediction system based on a hierarchical representation learning framework according to an embodiment of the present application;

[0085] Figure 4 is a hardware structure block diagram of a computer terminal of a drug-target association prediction method based on a hierarchical representation learning framework according to an embodiment of the present application. DETAILED DESCRIPTION

[0086] In order to make the objects, technical solutions and advantages of the present application clearer, the present application is further described in detail below with reference to the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and do not limit the present application.

[0087] According to an embodiment of the present application, method embodiments of a drug-target association prediction method based on a hierarchical representation learning framework are provided. It should be noted that the steps shown in the flowchart of the drawings can be executed in a computer system such as a group of computer executable instructions, and although a logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in a different order.

[0088] Figure 1 is a flowchart of a drug-target association prediction method based on a hierarchical representation learning framework according to an embodiment of the present application, as Figure 1 shown, the method comprises the following steps:

[0089] S101, in the constructed heterogeneous graph containing drugs, proteins, diseases and side effects, drug subgraphs and protein subgraphs are extracted according to different meta-paths, and the hierarchical feature learning framework MHGIN of dynamic neighborhood search is applied to the drug subgraph and the protein subgraph respectively to learn the drug and protein features;

[0090] In step S101 of the present application, a heterogeneous graph network containing multiple source biological entities is constructed, the nodes of which cover drugs (SMILES molecular representation), proteins (amino acid sequence features), diseases (MeSH subject heading annotation) and pathways (KEGG identifier association), and multiple edge types such as drug-target interaction, protein interaction and drug-disease association are defined through authoritative biological databases (ChEMBL, STRING, CTD) to form a complex biological heterogeneous graph network containing multiple types of nodes and multiple types of relationships;

[0091] The hierarchical feature learning framework of dynamic neighborhood search (Multi-Hop Graph Integration Network, MHGIN) of the application: it simultaneously captures the features of the graph data from the micro, meso and macro levels through a multi-channel fusion mechanism, uses a shallow network to extract local features, and a deep network to obtain global features through multi-hop convolution, fully captures long-range dependencies, and improves the understanding and analysis ability of the graph data;

[0092] In the graph data, long-range dependency refers to the interaction or correlation between nodes that are far apart. For example, in a biological system, the interaction between a drug molecule and a target protein may be influenced by surrounding environmental molecules and other indirectly related factors in the biological pathway. The relationship between these long-distance factors is the long-range dependency. Therefore, effectively capturing long-range dependencies is crucial for a comprehensive understanding of the behavior of complex systems.

[0093] Through the above embodiments, the application introduces a high-density information neighborhood search and a dynamic threshold screening mechanism to accurately identify high-density key nodes and reduce low-degree and low-density noise nodes.

[0094] S102, in each network layer of the hierarchical feature learning framework, high information density nodes are screened based on a fluctuation threshold, TOP-K neighbors are screened, and only high-density nodes screened by the threshold are retained for calculation; after reducing low-degree noise nodes, the topology connection is reconstructed using virtual edges, and all drug-protein pairs DPPs are traversed.

[0095] In the embodiments of the application, the virtual edge reconstruction strategy is combined to retain the long-range dependency path, solving the node misjudgment and connectivity destruction problems caused by static degree screening in traditional methods.

[0096] S103, the drug molecule features extracted by the dynamic neighborhood search framework MHGIN are tensor-level spliced with the protein semantic features generated by the denoising autoencoder DAE, the dimension is compressed through a learnable weight matrix, and a joint feature representation of the drug-protein pair DDP is formed.

[0097] S104, a multi-modal feature fusion strategy is adopted, a multi-head attention mechanism is used to assign dynamic weights to the multi-view features based on the drug-protein pair, the multi-view feature vectors are spliced, the key information is screened through the attention mechanism, and then input into a fully connected neural network. The correlation prediction value is output based on the fully connected neural network; wherein the multi-view includes a topology graph, a semantic graph and a collaboration graph, and the collaboration graph is obtained by averaging the sum of the feature representations of the topology graph and the semantic graph.

[0098] The application fuses the multi-order characteristics of topology graph, semantic graph and collaboration graph, realizes efficient integration of multi-view features through tensor-level splicing and attention mechanism, finally breaks through the dependence of shallow network on local neighborhood, improves the prediction accuracy and model robustness of implicit interaction in complex biological network, and comprehensively improves the modeling ability and prediction accuracy of long-range dependence relationship in complex biological network.

[0099] Based on the high information neighborhood search and low node reduction method, the application can accurately filter out high information density nodes and effectively reduce the noise interference caused by low nodes in sparse networks by dynamically generating fluctuation thresholds. Meanwhile, in order to maintain the topological connectivity of the network, the application reconstructs the virtual edge to ensure that the structure and information transmission of the network are not seriously affected after the low nodes are reduced, thereby improving the robustness of the model in sparse biological networks.

[0100] In addition, based on the dynamic neighborhood search hierarchical framework (MHGIN) of multi-level feature learning of complex biological network, the core is to solve the over-smoothing and long-range semantic association modeling defects of traditional graph neural network through hierarchical dynamic filtering mechanism and multi-level feature fusion strategy. Specifically, a dynamic neighborhood search module based on feature density and functional correlation is constructed at the bottom, a neighborhood weight is assigned through a learnable density perception function, low-density low-noise nodes are suppressed, and high-density key node aggregation is strengthened; a cross-layer attention mechanism is introduced at the middle layer, a multi-hop path semantic association matrix is constructed, long-range dependence is explicitly modeled through graph diffusion or path weight optimization, and the problem of long-distance signal dilution is solved; a hierarchical feature fusioner is designed at the top layer, and the neighborhood features of different hops and filtering weights are adaptively fused according to the semantic level, forming a multi-level feature representation containing local characteristics, mid-range association and long-range dependence.

[0101] Preferably, in step S102, the node degree is counted by a drug-target adjacency matrix A in the heterogeneous graph network , The node density is obtained by the number of connections, the degree of each node in each layer of the network is calculated, the threshold value is set according to the degree distribution characteristics, the nodes with a degree lower than the threshold value are selected and reduced, a high-density neighborhood information network is constructed, and the effective information in the sparse network is accurately preserved.

[0102] In the calculation of node degree and node density, the adjacency matrix of the original drug-protein heterogeneous graph network is A, wherein each element represents the connection relationship between node i and node j, the degree of each node in each layer of the network is calculated is the number of connections of node i with other nodes, and is represented as follows:

[0103]

[0104] wherein, is the total node degree in the network;

[0105] The node degree represents the number of connections between different nodes, and cannot measure the weight of the edge (such as the interaction strength, functional correlation and other key information), so the node density is introduced The average value of the neighborhood edge weight is calculated to reflect the connection quality and suppress the interference of noise on the attention mechanism and feature aggregation, thereby improving the long-range dependence modeling capability, and the node density is represented as:

[0106]

[0107] wherein, represents the connection weight between nodes i and j, and N(i) represents the neighborhood of node i.

[0108] Further, the step of setting a threshold according to the degree distribution characteristics and determining a dynamic threshold according to the inter-class variance includes:

[0109] First, draw a node degree histogram according to the calculated node degree, and calculate the probability distribution;

[0110] The node degree value is divided into L degrees, and the total node degree is N deg , wherein the maximum node degree is D, and the node degree range is [0, D];

[0111] wherein, the number of nodes in each node degree level is , the probability of each node degree level is calculated , and is represented as: ;

[0112] According to the data distribution and the experimental target, the threshold K is preliminarily determined as a percentile , represented as , wherein, represents the upward rounding function;

[0113] In addition, in order to avoid the limitations of the traditional fixed threshold, ensure that high-value nodes can be accurately screened and noise can be suppressed in different data sets and different network structures, thereby improving the generalization ability and robustness of the model, the dynamic fluctuation threshold mechanism is introduced to cope with the changes of different nodes in different data sets.

[0114] wherein, the fluctuation threshold provided in the embodiments of the present application aims to make the difference (inter-class variance) between nodes with high contact degree and nodes with low contact degree as large as possible, and the difference (intra-class variance) within each is as small as possible;

[0115] Therefore, the present invention divides nodes into two categories according to the threshold K:

[0116] Low-degree nodes : ;

[0117] Height Node : ;

[0118] Compute the cumulative probability and mean over two intervals:

[0119] Cumulative probability of low-degree node class Expressed as: ;

[0120] Cumulative probability of high-degree node class Expressed as: ;

[0121] Cumulative mean of low-degree node classes Expressed as: ;

[0122] Cumulative mean of high-degree node classes Expressed as: ;

[0123] Total mean drug-protein node degree Expressed as: ;

[0124] Calculate between-class variance Expressed as: ;

[0125] Finding the Fluctuation Threshold Make the difference between nodes with high and low connection degrees as large as possible, expressed as:

[0126]

[0127] Where, represents the probability of each node degree level, ε represents the degree level of each node, K represents the preliminary threshold, and D represents the maximum node degree; represents the between-class variance, represents the node degree;

[0128] Fluctuation threshold Adjustments are made based on changes in network structure, the density of node information, and the characteristics of the data set. When the network structure changes drastically, the fluctuation threshold will be appropriately adjusted to better eliminate redundant nodes and retain key nodes and long-range dependent paths.

[0129] Furthermore, in the embodiment of the present application, according to the fluctuation reduction threshold Filter out low-degree nodes with degrees below the threshold for elimination, quickly filter out obviously redundant nodes, select nodes with higher information density, and build a high-density information network;

[0130] In step 102 of the embodiment of the present application, the step of reconstructing the topological connection according to the virtual edge weight formula includes:

[0131] After the low-noise nodes are eliminated, they are connected to nodes with high information density in the heterogeneous graph network by adding virtual edges. To retain important topological information, it is expressed as: Where, represents the connection weight of the virtual edge, which is assumed to be a constant value and does not transmit information, but only transmits topological information; A represents the adjacency matrix; Represents the new adjacency matrix updated after adding virtual edges.

[0132] Furthermore, in the step of extracting features of different orders in the present application, a hierarchical feature learning framework (MHGIN) layer with dynamic neighborhood search is used to extract features of different orders;

[0133] Specifically, in step S101, for setting the order ; Obtain abstract features through the NGNN layer , expressed as: ,in, Indicates the The calculation function of the NGNN layer realizes the efficient modeling of the multi-level association of drug-protein bioinformatics network through hierarchical integration and dynamic screening of multi-level information; , the input feature matrix is , N is the number of samples, is the dimension of the input features, represents the set of real numbers, the adjacency matrix ;

[0134] Furthermore, after extracting features of different orders, global average pooling is performed to calculate the average pooling of each abstract feature matrix. Find the mean along the sample dimension and get , expressed as: , mean represents the mean operation, dim represents the dimension, , and then concatenate these mean vectors in the 0th dimension to get , the dimension of the output feature is .

[0135] Then, in the attention weight calculation, first pass through the first fully connected layer right The transformation is performed to obtain , which is expressed as:

[0136]

[0137] In the formula, The weight matrix of , T represents the transpose operation of the matrix, and the bias vector is , r represents the dimension reduction factor; is a rectified linear unit activation function, ;

[0138] Then, the second fully connected layer is used to transform , the weight matrix of is , and the bias vector is to obtain: , Finally, the Softmax function is used to normalize along the 0th dimension to obtain the attention weight matrix, which is expressed as:

[0139]

[0140] where i represents the index of the i-th view in the attention weighting, i = 1, 2,... ; j represents the index of the output feature dimension, j = 1, 2,... d out , k represents the index of the view feature, which is used for the normalization term in the denominator of Softmax, and the range is consistent with i; , d out represents the dimension of the output feature, represents the number of layers in the graph neural network, and for each column j, assigns each feature dimension indexed by i a probability distribution over different features or feature combinations as attention weights; exp represents the calculation of the exponential function; z ij represents the value of the i-th row and the j-th column, z kj represents the value of the k-th row and the j-th column; This allows the model to dynamically adjust the attention to different features or feature combinations according to these weights, highlighting important features and suppressing unimportant features, thereby better capturing key information in the data and improving the performance and expressiveness of the model;

[0141] In feature weighting and fusion, each abstract feature matrix is multiplied element-wise with the corresponding attention weight vector (the k-th row of ) to obtain the weighted feature matrix , which is expressed as: wherein, denotes element-wise multiplication, ; finally, all weighted feature matrices are added to obtain the final output Z, denoted as:

[0142]

[0143] wherein, k denotes a view index, k = 1,..., ; Z includes drug features and protein features .

[0144] In the encoding and decoding operations of the above-mentioned embodiment denoising autoencoder DAE, the DAE (denoising autoencoder) extracts features from multiple drug-related networks and protein-related networks, can embed high-dimensional original features into a low-dimensional space, remove redundant information, retain key features, improve the training efficiency and performance of the model; after these features are spliced into a large matrix representation, dimension reduction processing is performed by the DAE (denoising autoencoder) to obtain low-dimensional drug node features and protein node features represented by and respectively, and these features are collectively referred to as X.

[0145] In the above-mentioned embodiments of the present application, the drug embedding and target embedding obtained by the meta-path guided MHGIN and DAE are spliced in the column dimension respectively, denoted as:

[0146]

[0147]

[0148] In the formula, and denote the final drug features and protein features respectively.

[0149] Further, in step S104 of the embodiments of the present application, the construction process of the topological graph and the semantic graph includes the following steps:

[0150] All drug-protein pairs are traversed, and it is checked whether each drug-protein pair has a common drug or protein;

[0151] If there is a common drug or protein, the association value is set to 1 at the corresponding position of the adjacency matrix , otherwise 0;

[0152] The topological graph The structured relationship between drugs or proteins based on DPPs will be displayed, wherein each node in the topological graph represents a DPP, and each edge represents the association between DPPs.

[0153] The feature representation calculation formula based on the topological graph is:

[0154]

[0155] In the formula, represents the topological graph the feature representation at the i-th layer, which is based on the feature of the previous layer (i-1-th layer) is calculated, and more representative topological features are gradually extracted as the number of layers i increases; represents the rectified linear unit activation function, which is used to introduce a nonlinear factor into the model, because many interactions in actual biological relationships are nonlinear, and the model can learn more complex feature patterns through the function , and the expression ability of the model for the topological features of drug-protein pairs is enhanced; represents the normalization operation on the adjacency matrix ; is obtained by adding a self-loop (i.e., each node is connected to itself) to the original adjacency matrix ; adding a self-loop allows the node to consider its own feature information when updating the feature; is the degree matrix of , and the elements on the diagonal are the sum of the non-zero elements of the corresponding row or column in ; the power of is taken times, and then multiplied by and , which can normalize the adjacency information of the node, so that different nodes have comparability when aggregating the adjacency node information, which is helpful for the effective transmission and aggregation of information between nodes; represents the topological graph the trainable parameter matrix at the i-th layer, which performs linear transformation on the input feature , adjusts the weight and dimension of the feature, and enables the model to learn the complex relationship between different features; through the training process, the present application constantly optimizes these parameters to improve the learning ability of the model for the topological features of drug-protein pairs;

[0156] Similarly, the feature representation calculation formula of the semantic graph is:

[0157]

[0158] In the formula, represents the semantic graph the feature representation at the i-th layer; represents the rectified linear unit activation function; Represents the adjacency matrix Normalization operation; is the original adjacency matrix Add self-loop on the basis of yes The degree matrix of The sum of the non-zero elements in the corresponding rows or columns; Representing semantic graphs The trainable parameter matrix at layer i;

[0159] Furthermore, in the embodiment of the present disclosure, the feature representation calculation formula of the collaboration graph is:

[0160]

[0161] Where, Representing a collaboration diagram Feature representation at layer i; Represents a topology graph In the feature representation of layer i, Representing semantic graphs Feature representation at layer i;

[0162] In the calculation of the features of the collaboration graph in this embodiment, an average operation is used to assume that the information contained in the topological graph and the semantic graph is equally important for describing the features of the DPP (drug-protein pair) nodes.

[0163] Please refer to Figure 2 In step 104, a multi-head attention mechanism is used to assign dynamic weights to multi-view features based on drug-protein pairs, and the step of concatenating multi-view feature vectors includes the following steps:

[0164] S1041. Calculate the attention vector, expressed as:

[0165]

[0166] Where, represents the attention vector, Corresponding to the topological feature representation , semantic feature representation , collaborative feature representation Attention score; Represents the attention function, which is used to represent the topological features , semantic feature representation and collaborative feature representation , assign an attention score to each feature representation to measure their importance in representing the node;

[0167] For each drug-protein pair in the topological graph, semantic graph, and collaborative graph, a joint feature representation is generated , a query matrix is generated through a learnable linear mapping , a key matrix and a value matrix , wherein, , , is a weight matrix, denotes the real set, and N denotes the number of samples, denotes the dimension of the joint feature representation, and d k denotes the attention dimension;

[0168] Under the multi-head attention mechanism, the attention score is calculated and weightedly fused through combination of multiple Q, , , and is represented as:

[0169]

[0170] wherein, is a scaling factor, used for normalizing the dot product result of the query Q and the key vector;

[0171] S1042, a weight score is calculated, and the topological feature embedding is represented as:

[0172]

[0173] In the formula, W represents a learnable parameter matrix, used for linear transformation of the topological feature embedding, adjusting the dimension and weight of the feature, so that the model can learn the complex relationship between different features; b represents a learnable bias vector, adding a bias term to the result of the linear transformation to increase the flexibility of the model; q represents a learnable attention vector; denotes a hyperbolic tangent activation function, performing a nonlinear transformation on the result of , and then performing dot product operation with to obtain the weight score , which reflects the preliminary measure of the relative importance of the topological feature embedding when representing node i; the parameter T represents the transpose operation of the matrix, and the transpose operation plays an important role in adjusting the dimension of the matrix to meet the matrix multiplication rule in the calculation process of this attention mechanism, ensuring that the correct operation can be performed between the matrices and vectors, thereby realizing the weight calculation of different feature representations;

[0174] Similarly, the semantic feature embedding weight score of the topology feature and collaborative feature embedding weight score of the topology feature .

[0175] S1043, normalize the weight score, calculate the attention coefficient, denoted as:

[0176]

[0177] In the formula, denotes the attention coefficient of the topology feature, denotes the weight score of the topology feature, denotes the weight score of the semantic feature, denotes the weight score of the collaborative feature; the softmax(·) function normalizes , , all the attention coefficients to 1, and each attention coefficient is in the interval [0, 1], so that topology feature representation considering the relative importance of the three kinds of feature representations;

[0178] Similarly, the attention coefficient of the semantic feature representation and the attention coefficient of the collaborative feature representation are calculated.

[0179] S1044, using the calculated attention coefficients , and to weight and fuse the topology feature representation , the semantic feature representation and the collaborative feature representation , thereby obtaining the fused feature representation of node i. This weighted fusion method can reasonably combine different feature representations according to their importance, so that can more comprehensively and accurately represent the features of node i, providing more effective feature input for subsequent prediction of drug-target association scores, wherein .

[0180] Preferably, in step S104, the step of outputting the association prediction value based on the fully connected neural network comprises:

[0181] The obtained fused feature representation is input into a multi-layer fully connected network, and the fully connected neural network learns the complex relationship between the features and the drug-target association by adjusting the weights and biases in the network, and models the prediction task as a binary classification problem: When the correlation score output by the network is greater than the threshold value, it is determined that there is an interaction between the drug and the target;

[0182] When the correlation score output by the network is not greater than the threshold value, it is determined that there is no interaction.

[0183] When the correlation score output by the network is not greater than the threshold value, it is determined that there is no interaction.

[0184] Specifically, the prediction task of the embodiment of the present application is modeled as a binary classification problem, which means that the relationship between the drug-target is simply divided into two cases: “there is a correlation” and “there is no correlation”; in implementation, a threshold value is set, when the correlation score output by the network is greater than the threshold value, it is determined that there is an interaction between the drug and the target; when the score is not greater than the threshold value, it is determined that there is no interaction; this binary classification modeling method makes the problem more clear and easy to handle, which helps to quickly determine whether there is a potential interaction between the drug and the target, and provides an important reference for drug research and related research.

[0185] Figure 3 is a structural diagram of a drug-target correlation prediction system based on a hierarchical representation learning framework according to the embodiment of the present application;

[0186] As shown in Figure 3 , the correlation prediction system includes the following modules:

[0187] The feature learning module 201 is used to extract drug subgraphs and protein subgraphs according to different meta-paths in the constructed heterogeneous graph containing drugs, proteins, diseases and side effects, and apply a hierarchical feature learning framework MHGIN of dynamic neighborhood search to the drug subgraph and the protein subgraph respectively to learn drug and protein features;

[0188] The node reduction and reconstruction module 202 is used to filter high information density nodes based on a fluctuation threshold in each network of the hierarchical feature learning framework, filter TOP-K neighbors, and only keep the high density nodes after threshold filtering to participate in calculation; after reducing the low noise nodes, the topology connection is reconstructed by using virtual edges, and all drug-protein pairs DPPs are traversed;

[0189] The feature splicing module 203 is used to splice the drug molecular features extracted by the dynamic neighborhood search framework MHGIN and the protein semantic features generated by the denoising autoencoder DAE in a tensor level, compress the dimension through a learnable weight matrix, and form a joint feature representation of the drug-protein pair DDPs;

[0190] The correlation prediction module 204 is used to adopt a multi-modal feature fusion strategy, assign dynamic weights to multi-view features based on drug-protein pairs by using a multi-head attention mechanism, splice multi-view feature vectors, input a fully connected neural network after filtering key information by the attention mechanism, and output a correlation prediction value based on the fully connected neural network.

[0191] Figure 4 An internal structure diagram of a computer device in one embodiment is shown.

[0192] As Figure 4 shown, the computer device includes a processor, a memory, a network interface, an input device and a display screen connected through a system bus. The memory includes a non-volatile storage medium and an internal memory. The non-volatile storage medium of the computer device stores an operating system, and can also store a computer program which, when executed by the processor, can enable the processor to implement the drug-target association prediction method based on the hierarchical representation learning framework.

[0193] The computer program can also be stored in the internal memory, and when executed by the processor, can enable the processor to execute the drug-target association prediction method based on the hierarchical representation learning framework. The display screen of the computer device can be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device can be a touch layer overlaid on the display screen, or can be a key, trackball or touchpad arranged on the shell of the computer device, or can be an external keyboard, touchpad or mouse, etc.

[0194] Those skilled in the art can understand that Figure 4 the structure shown in the above is only a block diagram of part of the structure related to the scheme of the present application, and does not constitute a limitation on the computer device to which the scheme of the present application is applied. The specific computer device can include more or fewer components than those shown in the figure, or combine certain components, or have a different arrangement of components.

[0195] In one embodiment, the drug-target association prediction system based on the hierarchical representation learning framework provided by the present application can be implemented in the form of a computer program, which can run on a computer device as Figure 4 shown. The memory of the computer device can store various program modules constituting the drug-target association prediction system based on the hierarchical representation learning framework. The computer program constituted by the various program modules enables the processor to execute the steps in the drug-target association prediction method based on the hierarchical representation learning framework of each embodiment of the present application described in the specification.

[0196] It should be understood that, although the steps in the flowcharts of the embodiments of the present application are shown in a certain order according to the arrows, the steps are not necessarily executed in the order of the arrows. Unless otherwise specified herein, the execution of the steps is not strictly limited in order, and the steps can be executed in other orders. Moreover, at least some of the steps in the embodiments can include a plurality of sub-steps or a plurality of stages, which are not necessarily executed at the same time, but can be executed at different times, and the execution order of the sub-steps or stages is not necessarily sequential, but can be round-robin or alternately executed with at least some of the other steps or sub-steps or stages of the other steps.

[0197] It can be understood by those skilled in the art that all or part of the processes in the above-mentioned embodiments can be completed by a computer program instructing related hardware, and the program can be stored in a non-volatile computer readable storage medium. When the program is executed, it can include the processes of the above-mentioned embodiments of the methods. Any reference to memory, storage, database or other medium used in the embodiments provided in the present application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM) or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. As an illustration but not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), double data rate SDRAM (DDR SDRAM), enhanced SDRAM (ESDRAM), synchronous link (Synchlink) DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0198] The technical features of the above-mentioned embodiments can be combined in any way. In order to make the description concise, all possible combinations of the technical features in the above-mentioned embodiments are not described, but as long as the combination of the technical features does not exist, it should be considered as the scope of the present application.

[0199] The above embodiments only express several implementation manners of the present application, which are described in a more specific and detailed manner, but should not be understood as a limitation on the patent scope of the present application. It should be noted that, for ordinary skilled persons in the art, several modifications and improvements can be made without departing from the concept of the present application, which are all within the protection scope of the present application. Therefore, the protection scope of the patent of the present application should be subject to the appended claims.

[0200] The above merely describes the preferred embodiments of the present application and should not be used to limit the present application. Any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A drug-target association prediction method based on a hierarchical representation learning framework, characterized by: The following steps are involved: S101. Extract drug subgraphs and protein subgraphs based on different meta-paths from a constructed heterogeneous graph containing drugs, proteins, diseases, and side effects. Apply the hierarchical feature learning framework MHGIN with dynamic neighborhood screening to the drug subgraphs and protein subgraphs to learn drug and protein features. The underlying structure of the hierarchical feature learning framework MHGIN is based on a dynamic neighborhood screening module based on feature density and functional correlation. It assigns neighborhood weights through a learnable density-aware function, suppresses low-degree and low-density noise nodes, and strengthens the aggregation of high-density key nodes. The hierarchical feature learning framework MHGIN introduces a cross-layer attention mechanism in the middle layer to construct a multi-hop path semantic association matrix and explicitly model long-range dependencies through graph diffusion or path weight optimization. The hierarchical feature fuser, the top-level design of the hierarchical feature learning framework MHGIN, adaptively fuses neighborhood features of different hop counts and filtering weights according to the semantic level to form a multi-level feature representation that includes local characteristics, medium-range correlations, and long-range dependencies. S102. In the underlying network of the hierarchical feature learning framework, high-information-density nodes are screened based on the fluctuation threshold, and the TOP-K neighbors are screened. Only high-density nodes that have passed the threshold screening are retained for calculation; after eliminating low-noise nodes, the topological connections are reconstructed using virtual edges to traverse all drug-protein pairs (DPPs); S103, tensor-level concatenation of drug molecular features extracted by the dynamic neighborhood screening framework MHGIN and protein semantic features generated by the denoising autoencoder DAE is performed, and the dimensions are compressed by a learnable weight matrix to form a joint feature representation of drug-protein pairs (DDPs); S104. A multimodal feature fusion strategy is adopted, and a multi-head attention mechanism is used to assign dynamic weights to multi-view features based on drug-protein pairs. The multi-view feature vectors are spliced, and after the key information is filtered by the attention mechanism, the information is input into a fully connected neural network, and the association prediction value is output based on the fully connected neural network.

2. The drug-target association prediction method based on a hierarchical representation learning framework according to claim 1, characterized in that: In step S101, for setting the order =1, 2, ... ; Obtain abstract features through the NGNN layer , expressed as: ;in, Indicates the The calculation function of the NGNN layer, , the input feature matrix is , N is the number of samples, is the dimension of the input features, Represents the set of real numbers; adjacency matrix ; For each abstract feature matrix Find the mean along the sample dimension and get , expressed as: ; In the formula, mean represents the mean operation, dim represents the dimension, , and then concatenate the mean vectors in dimension 0 to get , the dimension of the output feature is d out ; In the attention weight calculation, first pass through the first fully connected layer right Transform and get , expressed as: ; Wherein, the weight matrix of FC1 is , T represents the transpose operation of the matrix, and the bias vector is , r represents the dimensionality reduction factor; ReLU(x)=max(0,x) is the rectified linear unit activation function, ; Then Z1 is transformed through the second fully connected layer FC2, and the weight matrix of FC2 is , the bias vector is ,get: , Finally, the Softmax function is used to normalize Z2 along the 0th dimension to obtain the attention weight matrix, which is expressed as: ； Where i represents the index of the i-th view in the attention weighting, i=1,2,... ; j represents the dimension index of the output feature, j=1,2,...d out ; k represents the index of the view feature, which is used for the normalization term of the Softmax denominator and has the same range as i; , d out Represents the dimension of the output feature, and for each column j there is , assigns a probability distribution on different features or feature combinations as attention weights to each feature dimension indexed by i; exp represents the calculation of the exponent; z ij represents the value of row i and column j, z kj Represents the value of the kth row and jth column; In feature weighting and fusion, each abstract feature matrix And the corresponding attention weight vector Element-wise multiplication, where Represents the weight vector corresponding to the kth view of the attention weight matrix, and obtains the weighted feature matrix ; Add all weighted feature matrices to get the final output Z, which includes drug feature Z drug and protein feature Z protein .

3. The drug-target association prediction method based on the hierarchical representation learning framework according to claim 2 is characterized in that: In step S102, the node degree d is counted by the adjacency matrix A of the drug-protein network in the heterogeneous graph network. i , d i It is used to reflect the number of node connections, obtain the node density through the relationship between the number of connections, calculate the degree of each node in each layer of the network, set the threshold according to the statistical distribution characteristics of the degree, and determine the dynamic threshold based on the inter-class variance.

4. The drug-target association prediction method based on a hierarchical representation learning framework according to claim 3, characterized in that: The steps of setting a threshold value according to the degree statistical distribution characteristics and determining a dynamic threshold value according to the inter-class variance include: There are L degree levels for statistical node degree values, and the total node degree is N deg , where the maximum node degree is D and the node degree range is [0, D]; The number of nodes of degree level ε at each node , calculate the probability of each node degree level , expressed as: ; Preliminary determination of the percentile of the threshold K , expressed as ; in, represents the ceiling function; The nodes are divided into two categories according to the threshold K: Low-degree nodes : ; Height Node : ; Compute the cumulative probability and mean over two intervals: Cumulative probability of low-degree node class Expressed as: ; Cumulative probability of high-degree node class Expressed as: ; Cumulative mean of low-degree node classes Expressed as: ; Cumulative mean of high-degree node classes Expressed as: ; Total mean drug-protein node degree Expressed as: ; Calculate between-class variance Expressed as: ; Finding the Fluctuation Threshold Make the difference between nodes with high and low connection degrees as large as possible, expressed as: ； Where, represents the probability of each node degree level, ε represents the degree level of each node, K represents the preliminary threshold, and D represents the maximum node degree; represents the between-class variance, Represents the node degree.

5. The drug-target association prediction method based on a hierarchical representation learning framework according to claim 4, characterized in that: In step 102, the step of reconstructing the topological connection using the virtual edge includes: After the low-noise nodes are eliminated, they are connected to nodes with high information density in the heterogeneous graph network by adding virtual edges. To preserve important topological information.

6. The drug-target association prediction method based on a hierarchical representation learning framework according to claim 5, characterized in that: In step S104, the multi-view includes a topological map, a semantic map, and a collaboration map. The collaboration map is obtained by adding and averaging the feature representations of the topological map and the semantic map. The feature representation calculation formula of the collaboration map is: ； Where, Represents the collaboration graph G c Feature representation at layer i; Represents the topology graph G t In the feature representation of layer i, Represents the semantic graph G s Feature representation at layer i.

7. The drug-target association prediction method based on a hierarchical representation learning framework according to claim 6, characterized in that: In step 104, the multi-head attention mechanism is used to assign dynamic weights to the multi-view features based on the drug-protein pair, and the step of concatenating the multi-view feature vectors includes: Calculate the attention vector, expressed as: ； Where, represents the attention vector, Corresponding to the topological feature representation , semantic feature representation , Collaborative Feature Representation Att(·) represents the attention function; Joint feature representation of each drug-protein pair in topological graph, semantic graph, and collaboration graph , generate the query matrix through a learnable linear mapping , key matrix Sum Matrix ,in, 、 、 is the weight matrix, represents a set of real numbers, N represents the number of samples, represents the dimension of the joint feature representation, d k represents the attention dimension; Under the multi-head attention mechanism, by 、 The combined attention scores are calculated and weighted fused, expressed as: ； in, is the scaling factor used to compare the query Q with the key The vector dot product result is normalized; Calculate weight scores and topological feature embedding The weight score is expressed as: ； Where W represents a learnable parameter matrix, which is used to perform linear transformation on topological feature embedding and adjust the dimension and weight of features; b represents a learnable bias vector; q represents a learnable attention vector; represents the hyperbolic tangent activation function, The result is then transformed nonlinearly and then Do the dot product operation to get the weight score ; Parameter T represents the transpose operation of the matrix; Similarly, computing semantic feature embedding Weight score and collaborative feature embedding Weight score ; Normalize the weight scores and calculate the attention coefficient, which is expressed as: ； Where, represents the attention coefficient of topological features, represents the weight score of the topological feature, represents the weight score of the semantic feature, represents the weight score of the collaborative feature; The function will 、 、 Normalize them together so that the sum of all attention coefficients is 1 and each attention coefficient is in the interval [0,1]; Similarly, the attention coefficient of semantic feature representation is calculated and the attention coefficient of collaborative feature representation ; Using the attention coefficient 、 and Representation of topological features , semantic feature representation and collaborative feature representation Perform weighted fusion to obtain the fused feature representation of node i.

8. The drug-target association prediction method based on a hierarchical representation learning framework according to claim 7, characterized in that: In step S104, the step of outputting the association prediction value based on the fully connected neural network includes: The obtained fused feature representation is input into a multi-layer fully connected network. By adjusting the weights and biases in the network, the fully connected neural network learns the complex relationship between features and drug-target associations and models the prediction task as a binary classification problem: When the association score output by the network is greater than the threshold, it is determined that there is an interaction between the drug and the target; When the correlation score output by the network is not greater than the threshold, it is determined that there is no interaction.

Images