A drug resistance prediction system and method based on tensor enhancement graph similarity

Abstract

This invention discloses a drug resistance prediction system and method based on tensor-enhanced graph similarity, belonging to the field of computer-aided drug detection technology. It includes: an acquisition module to obtain drug compound and three-dimensional structural data; a preprocessing module to construct a graph structure; a node embedding learning module to learn node features through graph convolution and graph self-attention mechanisms; a graph interaction modeling module to perform inter-graph node interactions based on t-products; a similarity matrix learning module to predict graph structure similarity scores; and a result output module to output drug resistance results based on the scores. This invention comprehensively captures drug structural features and improves prediction accuracy through multi-view tensor modeling, node embedding learning, and other techniques; improves computational efficiency through end-to-end joint training and multi-scale convolution; enhances model generalization ability through preprocessing and graph structure construction starting from the original data; and provides rich information for drug resistance assessment through a multi-semantic similarity matrix.

Description

Technical Field

[0001] This invention relates to the field of computer-aided drug detection technology, and more specifically to an intelligent prediction system and method for antiviral drug resistance based on tensor-enhanced graph similarity calculation. Background Technology

[0002] In recent years, Graph Neural Networks (GNNs) have been increasingly incorporated into tasks such as molecular similarity calculation and molecular activity prediction, overcoming the limitations of traditional molecular fingerprinting methods in terms of structural representation capabilities and prediction accuracy, and becoming one of the important research directions in the field of drug discovery. Compared to description methods based on SMILES or molecular fingerprints, GNNs can more naturally model the topological relationships and complex interactions between atoms in molecular structures, demonstrating excellent performance in tasks such as drug property prediction, target identification, and virtual screening.

[0003] However, when applying graph neural networks to antiviral drugs, especially in the scenario of predicting drug resistance in viruses with high mutation rates such as HIV, existing methods still have the following key problems, which seriously limit their practicality and generalization ability:

[0004] First, existing methods fail to adequately model the three-dimensional spatial complementarity of HIV protease targets. In actual molecule-target interactions, the degree of spatial matching between the molecule's conformation and the target's active site plays a decisive role in drug efficacy. Traditional two-dimensional graph modeling methods often neglect this three-dimensional structural docking relationship, leading to significant differences in drug efficacy even when molecular topologies are similar, due to conformational differences. This affects the model's accurate assessment of drug activity and resistance.

[0005] Secondly, existing graph neural networks have shortcomings in modeling molecular functional groups, failing to effectively capture key pharmacophore features such as hydrogen bond donors / receptors. Functional groups play a central role in the binding process between molecules and viral targets, especially non-covalent interactions such as hydrogen bonds, which are crucial for the stability of molecule-target complexes. However, current graph neural network structures generally lack explicit modeling of functional group semantics, making it difficult to accurately capture molecular fragments that drive drug efficacy, thus affecting the model's predictive ability and reliability.

[0006] Third, existing methods also have significant shortcomings in target-specific modeling. General graph similarity measurement methods fail to adequately adapt to the dynamic binding pocket characteristics of targets such as HIV proteases. When applied across different targets (such as HIV-PR and HIV-RT), the early enrichment capacity (EF1%) of existing models decreases significantly, limiting the practicality and broad applicability of the models in screening multi-target synergistic antiviral drugs.

[0007] Furthermore, traditional methods generally lack a mechanism for focusing on active substructures, and usually treat all substructures in a molecule equally, ignoring the preservation of key structural motifs such as protease hinges. This leads to the overestimation of the similarity of some molecules containing invalid or redundant fragments (such as flexible long chains), increasing the probability of screening false candidates and reducing the efficiency and accuracy of drug screening.

[0008] In practical applications, there is still an imbalance between efficiency and accuracy. While molecular docking-based methods have high prediction accuracy, they are computationally expensive, and the docking analysis of a single molecule is time-consuming, making it difficult to support high-throughput screening of large-scale molecular libraries. On the other hand, traditional methods such as molecular fingerprinting are computationally efficient, but they cannot fully express the three-dimensional structure and functional group characteristics of complex molecules, resulting in insufficient prediction accuracy and making it difficult to meet the needs of predicting complex viral drug resistance.

[0009] In summary, how to provide an intelligent prediction method and system for antiviral drug resistance that integrates strong structural feature modeling capabilities, target adaptability, and supports efficient graph similarity measurement and multi-scale feature extraction has become a technical problem that urgently needs to be solved in this field. Summary of the Invention

[0010] In view of this, the present invention provides a drug resistance prediction system and method based on tensor-enhanced graph similarity. Through four stages, namely node embedding learning, graph interaction modeling, similarity matrix alignment and similarity matrix learning, it achieves accurate characterization of three-dimensional pharmacodynamic features, targeted focusing of active substructures, enhanced adaptability to drug resistance mutations, efficient screening of target specificity, and rapid retrieval of large-scale libraries.

[0011] To achieve the above objectives, the present invention adopts the following technical solution:

[0012] On one hand, the present invention provides a drug resistance prediction system based on tensor-enhanced graph similarity, comprising:

[0013] The acquisition module is used to acquire compound and three-dimensional structural data of antiviral drugs;

[0014] The preprocessing module is used to preprocess the compound and three-dimensional structural data of the antiviral drug to construct a graph structure.

[0015] The node embedding learning module is used to learn the initial node features in the graph structure through graph convolution and graph self-attention mechanisms to obtain the node embedding matrix;

[0016] The graph interaction modeling module is used to construct a multi-view tensor based on the node embedding matrix using t-product, and to perform inter-graph node interaction through tensor quantum space learning to obtain a similarity matrix of multiple semantic perspectives.

[0017] A similarity matrix learning module is used to predict the similarity score of the graph structure based on the similarity matrix;

[0018] The result output module is used to output the drug resistance of the antiviral drug based on the similarity score.

[0019] Preferably, the node embedding learning module includes:

[0020] The residual graph convolutional unit is used to generate a local embedding matrix of a node by aggregating the features of the node and its neighbors.

[0021] The graph attention unit is used to capture the dependencies between nodes through a multi-head self-attention mechanism and combine it with the local embedding matrix to obtain the node embedding matrix.

[0022] Preferably, the graph interaction modeling module includes:

[0023] A node embedding construction unit is used to tensor the node features corresponding to the graph pair data into a tensor based on the t-product operation;

[0024] The Zhang quantum space learning unit is used to constrain the tensor based on the objective function to obtain a self-expressive tensor, and to obtain a similarity matrix from multiple semantic perspectives by accumulating the self-expressive tensors of each dimension.

[0025] Preferably, the graph interaction modeling module achieves end-to-end joint training through an embedded optimization layer, including:

[0026] By introducing auxiliary variable Z and Lagrange multipliers G, an augmented Lagrange function is constructed.

[0027] The optimization problem is expanded into a network structure with a fixed number of layers using the ADMM alternating update method, and an embedded optimization sub-network is constructed.

[0028] Preferably, the similarity matrix learning module includes:

[0029] A multi-scale cross-node similarity coding layer is used to perform multi-scale convolution on the similarity matrix using cross-shape filters of various sizes to obtain feature maps at different scales, and then fuses the feature maps using a weighted average method to obtain a fused feature map.

[0030] A node similarity learning layer is used to aggregate cross-node information by employing one-dimensional convolution along the node dimension in the fused feature map, generating node-level similarity embeddings;

[0031] A fully connected layer is used to embed the node similarity into a final graph similarity score.

[0032] Preferably, the system further includes:

[0033] The semantic alignment module is used to perform feature transformation and projection on the similarity matrices of multiple semantic perspectives output by the graph interaction modeling module, so as to unify the similarity expression of the similarity matrices.

[0034] On the other hand, the present invention provides a drug resistance prediction method based on tensor enhancement graph similarity, comprising the following steps:

[0035] To obtain compound and three-dimensional structural data of antiviral drugs;

[0036] The compound and three-dimensional structural data of the antiviral drug are preprocessed to construct a graphical structure;

[0037] The node embedding matrix is ​​obtained by learning the initial node features in the graph structure through graph convolution and graph self-attention mechanisms;

[0038] Based on t-product, a multi-view tensor is constructed from the node embedding matrix, and the interaction between nodes in the graph is performed through tensor quantum space learning to obtain a similarity matrix from multiple semantic perspectives.

[0039] Predict the similarity score of the graph structure based on the similarity matrix;

[0040] The drug resistance of the antiviral drug is output based on the similarity score.

[0041] As can be seen from the above technical solution, compared with the prior art, this invention discloses a drug resistance prediction system and method based on tensor-enhanced graph similarity. Firstly, by employing multi-view tensor modeling, node embedding learning, and semantic alignment, it comprehensively captures drug structural features from multiple perspectives, improving prediction accuracy. Secondly, end-to-end joint training and multi-scale cross-node similarity encoding improve computational efficiency and accelerate model training and prediction speed. Furthermore, preprocessing and graph structure construction starting from the original data, along with the combination of graph convolution and graph self-attention mechanisms, enhance the model's generalization ability to different types of drugs, enabling it to maintain good prediction performance across different datasets and application scenarios. Simultaneously, the multi-semantic similarity matrix provides richer information for drug resistance assessment, contributing to a deeper understanding of drug resistance mechanisms. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on the provided drawings without creative effort.

[0043] Figure 1 This is a schematic diagram of the overall structure of the present invention.

[0044] Figure 2 A detailed structural diagram is provided for this invention.

[0045] Figure 3 This is a schematic diagram of another embodiment of the present invention.

[0046] Figure 4 This is a schematic diagram of the process provided by the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0048] This invention discloses one aspect: a drug resistance prediction system based on tensor enhancement graph similarity, such as... Figure 1-2 As shown, it includes:

[0049] The acquisition module is used to acquire compound and three-dimensional structural data of antiviral drugs.

[0050] The preprocessing module is used to preprocess the compound and 3D structural data of antiviral drugs to construct graph structures. The chemical molecular graphs of the acquired antiviral drug compound and 3D structural data can be represented in SMILES format (Simplified Linear Representation of Molecular Input). Data preprocessing requires converting the molecular structure into a graph form, where nodes represent atoms and edges represent chemical bonds. Chemical computing libraries such as RDKit or Deep Chem are used to read and process these molecular graphs.

[0051] The node embedding learning module is used to learn the node embedding matrix from the initial node features in the graph structure through graph convolution and graph self-attention mechanisms. A pharmacophore attention mechanism is introduced in the node embedding learning stage, which increases the embedding weight of key active groups (such as the hydroxyethyl isosteric group of HIV protease inhibitors) by 3-5 times, significantly reducing interference from invalid fragments and lowering the false positive rate.

[0052] The graph interaction modeling module is used to construct a multi-view tensor based on the node embedding matrix using t-product, and to perform inter-graph node interactions through tensor quantum space learning, thereby obtaining similarity matrices from multiple semantic perspectives. Specifically, based on the acquired node embedding data, a multi-view tensor is constructed based on t-product, and inter-graph node interactions are performed through tensor quantum space learning. After low-rank constraints and graph regularization, not only is information within a single graph fused, but also multiple semantic perspective similarity matrices characterizing cross-graph interactions between two graphs can be obtained.

[0053] The similarity matrix learning module is used to predict the similarity score of a structure based on the similarity matrix.

[0054] The results output module is used to output antiviral drug resistance prediction results based on similarity scores. The prediction results include suggestions for optimizing anti-HIV activity: when the similarity is >0.7 and contains a "protease_hinge" substructure, resistance mutation experiments are recommended; molecules with similarity between 0.5 and 0.7 are labeled as "backbone transition candidates," providing a library of side chain substitution schemes. The output results are visualized, including molecular alignment overlay maps and pharmacophore matching heatmaps.

[0055] Specifically, the node embedding learning module includes:

[0056] The residual graph convolutional unit generates a local embedding matrix for a node by aggregating features of the node and its neighbors. Residual connections are introduced to avoid over-smoothing. The output local embedding matrix is ​​represented as follows:

[0057] ;

[0058] in, This represents the adjacency matrix with added self-loops (A is the original adjacency matrix, I is the adjacency matrix with added self-loops). N (where N is the identity matrix), and the node's own characteristics are preserved by introducing self-loops; Symmetric normalization was achieved. For degree matrix, This eliminates the bias in feature propagation caused by differences in node degree. The node feature matrix of layer l-1 is obtained through a trainable weight matrix. Complete the feature space projection and pass it through a nonlinear activation function. Introducing model expressive power; the final residual term H (l-1) By preserving original feature information through cross-layer connections, the vanishing gradient problem is effectively mitigated. This leads to the synergistic effect of neighborhood feature aggregation, nonlinear transformation, and residual learning, achieving efficient representation learning for graph-structured data.

[0059] The graph attention unit is used to capture the dependencies between nodes through a multi-head self-attention mechanism, and combines the local embedding matrix to obtain the node embedding matrix, which is represented as:

[0060] ;

[0061] in, , , Let represent the query, key, and value matrices of the m-th attention head, respectively. d k is the dimension scaling factor for the key vector (used for stable gradient calculation). The learnable structural correlation coefficient. The degree matrix bias term (encoding the connection strength between nodes) is based on the graph topology. The Softmax function normalizes the attention scores along the row direction, and the final output is obtained by fusing node feature similarity (from...). Representation) and prior information of graph structure (from rD) g Feature reconstruction is achieved by introducing a weighted vector. This computation mechanism enhances the adaptability of attention weights to network topology by explicitly incorporating spatial constraints of the graph.

[0062] Furthermore, the graph interaction modeling module includes:

[0063] Node embedding constructs are used to tensor the node features corresponding to graph pairs based on the t-product operation; for example, the node embedding matrices of graphs A and B are as follows: , , where N A With N B ...

[0064] ;

[0065] In this tensor, the first frontal slice corresponds to the information in Figure A, and the second frontal slice corresponds to the information in Figure B.

[0066] The Zhang quantum space learning unit is used to constrain tensors based on an objective function to obtain self-expressive tensors, and by accumulating the self-expressive tensors of each dimension, a similarity matrix from multiple semantic perspectives is obtained. Specifically, tensors... Self-expression tensors can be used Approximate reconstruction, i.e.:

[0067] ;

[0068] in Defined as t-product:

[0069] ;

[0070] Expand tensor C in the third dimension;

[0071] express The block cyclic matrix;

[0072] Reconstruct the product result into the original tensor format.

[0073] The constructed self-expressive tensor C contains data from two horizontal slices, where:

[0074] It can mainly reflect the internal structure of Figure A and the interaction between Figure A and Figure B;

[0075] This reflects the interaction within Figure B and between Figure B and Figure A.

[0076] To ensure that the self-expressive tensor C simultaneously possesses global low-rank properties and preserves local graph structure, the objective function is defined as follows:

[0077] ;

[0078] in, Denotes the Frobenius norm; The tensor nuclear norm (TNN) is used to force C to have a low-rank structure. ,in This represents the i-th horizontal slice data in the third dimension of the three-dimensional tensor after performing a Fourier transform on C; This represents the v-th horizontal slice of data; Let v be the graph Laplacian matrix corresponding to the viewpoint v, which is constructed based on the similarity between node embeddings;

[0079] Parameter θ and To control the weights of low-rank and graph regularization terms.

[0080] By performing absolute value operations on each horizontal slice and then transposing and summing them:

[0081] ;

[0082] This results in a similarity matrix that incorporates the interaction information between graph A and graph B. At the same time, each horizontal slice of data is retained as a representation from different semantic perspectives, namely: This forms a similarity matrix from multiple semantic perspectives.

[0083] Furthermore, to enable end-to-end joint training of the aforementioned optimization problem with the neural network, this invention achieves end-to-end joint training by embedding an optimization layer within the graph interaction modeling module, including:

[0084] By introducing auxiliary variable Z and Lagrange multipliers G, an augmented Lagrange function is constructed.

[0085] The optimization problem is expanded into a network structure with a fixed number of layers using the ADMM alternating update method, and an embedded optimization sub-network is constructed.

[0086] The specific steps are as follows:

[0087] By introducing an auxiliary variable Z (used to replace C in solving for the TNN term) and a Lagrange multiplier G, we construct the augmented Lagrange function:

[0088] ;

[0089] Where μ is the penalty parameter, This indicates the inner product operation.

[0090] The above optimization problem is expanded into a fixed-layer network structure using the ADMM alternating update method. Each layer t includes the following update steps:

[0091] During forward propagation, the Singular Value Thresholding (SVT) method is used to update each Fourier domain frontal slice:

[0092] ;

[0093] Z is obtained after inverse FFT t+1 .

[0094] Solve the following equation:

[0095] ;

[0096] This problem can be solved by decomposing each frontal slice in the FFT domain to obtain a closed or iterative update formula, and ensure that the update result is consistent with the output of the forward pass layer.

[0097] ;

[0098] By fixing the above update process and expanding it into T layers, all parameters of each layer can participate in backpropagation, forming a trainable embedded optimization subnetwork.

[0099] This invention employs a dynamic weight adjustment strategy to optimize the graph interaction modeling parameters for HIV protease targets, thereby improving the early enrichment rate (EF1%).

[0100] Furthermore, the similarity matrix learning module includes:

[0101] A multi-scale cross-node similarity coding layer is used to perform multi-scale convolution on the similarity matrix using cross-shape filters of various sizes to obtain feature maps at different scales. These feature maps are then fused using a weighted average method to obtain a fused feature map. Specifically, cross-shape filters of various sizes are designed. For example:

[0102] 3×3 cross filter Only the middle row and middle column are valid;

[0103] 5×5 cross filter ;

[0104] 7×7 cross filter ;

[0105] The formula for each filter is as follows (taking 3×3 as an example):

[0106] Among them, w ij It is a filter Learnable weight parameters located at the central cross position;

[0107] Perform multi-scale convolutions on the similarity matrix respectively:

[0108] ;

[0109] in, Representing different sizes, the output feature map shape is as follows

[0110] C s This represents the number of output channels at the corresponding scale.

[0111] Feature maps generated at different scales are fused into a unified representation through a weighted average:

[0112] ;

[0113] in These are the trainable weight parameters.

[0114] The node similarity learning layer is used to fuse feature maps by aggregating cross-node information through one-dimensional convolution along the node dimension to generate node-level similarity embeddings.

[0115] The fully connected layer maps node similarity embeddings to the final graph similarity score; specifically, the input is the fused multi-scale CSL layer output. ,in This represents the total number of channels across all scales.

[0116] 1D convolution along the node dimension (usually the row or column direction) is used to aggregate cross-node information.

[0117] For each node, its corresponding row (or column) is treated as a sequence, and a 1D convolution filter F is applied. NSL :

[0118] ;

[0119] in, For node-level similarity embedding, Let represent the activation function, and b be the bias term. Here, 1D convolution will capture the local aggregated features of each node at different locations (across nodes). , is a one-dimensional convolutional filter; k is the width of the convolutional kernel (sliding window size), which controls the coverage of the local neighborhood. The number of input channels corresponds to the multi-scale convolutional features. The total number of channels, where d is the number of output channels (i.e., the number of filter banks), determines the final node embedding. Dimensions.

[0120] In another embodiment, to fully utilize similarity matrices from different perspectives and ensure their consistency within a common semantic space, such as... Figure 3 As shown, this invention introduces a semantic alignment module to perform feature transformation and projection on the similarity matrices from multiple semantic perspectives output by the graph interaction modeling module, unifying the similarity expression of the similarity matrix. This embodiment integrates mutation sensitivity coefficients through the similarity matrix alignment stage to accurately predict the impact of common drug resistance mutations such as V82A and I84V on molecular activity. Specifically, it includes:

[0121] For each viewpoint v, set a trainable linear mapping matrix. The original similarity matrix Mapping to the public semantic space: ,in, This represents the similarity matrix after semantic alignment.

[0122] To ensure consistency in alignment results across different viewpoints, the following alignment loss is designed:

[0123] ;

[0124] The aim is to minimize the differences between perspectives, thereby establishing a consistent expression of similarity within a common semantic space.

[0125] For multiple aligned viewpoint matrices, a unified similarity matrix is ​​obtained through weighted fusion:

[0126] ;

[0127] in These are learnable weight parameters used to reflect the importance of each semantic perspective.

[0128] On the other hand, this invention provides a drug resistance prediction method based on tensor-enhanced graph similarity, such as... Figure 4 As shown, it includes the following steps:

[0129] To obtain compound and three-dimensional structural data of antiviral drugs;

[0130] Preprocessing of antiviral drug compound and three-dimensional structure data to construct graph structures;

[0131] The node embedding matrix is ​​obtained by learning the initial node features in the graph structure through graph convolution and graph self-attention mechanisms.

[0132] Based on t-product, a multi-view tensor is constructed from the node embedding matrix, and the interaction between nodes in the graph is performed through tensor quantum space learning to obtain a similarity matrix from multiple semantic perspectives.

[0133] Predict the similarity score of the graph structure based on the similarity matrix;

[0134] Antiviral drug resistance is output based on similarity scores.

[0135] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on its differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the apparatus disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple; relevant parts can be referred to the method section.

[0136] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A drug resistance prediction system based on tensor-enhanced graph similarity, characterized in that, include: The acquisition module is used to acquire compound and three-dimensional structural data of antiviral drugs; The preprocessing module is used to preprocess the compound and three-dimensional structural data of the antiviral drug to construct a graph structure. The node embedding learning module is used to learn the initial node features in the graph structure through graph convolution and graph self-attention mechanisms to obtain the node embedding matrix; an attention mechanism is introduced to weight the active groups during the node embedding learning stage; The graph interaction modeling module is used to construct a multi-view tensor based on the node embedding matrix using t-product, and to perform inter-graph node interaction through tensor quantum space learning to obtain a similarity matrix of multiple semantic perspectives. A similarity matrix learning module is used to predict the similarity score of the graph structure based on the similarity matrix; The result output module is used to output the drug resistance of the antiviral drug based on the similarity score; The node embedding learning module includes: The residual graph convolutional unit is used to generate a local embedding matrix for a node by aggregating the features of the node and its neighbors; the output local embedding matrix is ​​represented as: ； in, This represents the adjacency matrix with added self-loops, where A is the original adjacency matrix, and I... N It is an N-dimensional identity matrix; For symmetric normalization, For degree matrix, ; This is the feature matrix of the nodes in the (l-1)th layer. For trainable weight matrix, H is a non-linear activation function. (l-1) For residual terms; The graph attention unit is used to capture the dependencies between nodes through a multi-head self-attention mechanism and combine it with the local embedding matrix to obtain the node embedding matrix; the node embedding matrix is ​​represented as: ； in, , , Let d represent the query, key, and value matrix of the m-th attention head, respectively. k is the scaling factor for the dimension of the key vector. The learnable structural correlation coefficient. For the degree matrix bias term based on graph topology; The graph interaction modeling module includes: Node embedding construction unit, used to tensor the node features corresponding to nodes between graphs into tensors based on t-product operation; The Zhang quantum space learning unit is used to constrain the tensor based on the objective function to obtain a self-expressive tensor, and to obtain a similarity matrix from multiple semantic perspectives by accumulating the self-expressive tensors of each dimension.

2. The drug resistance prediction system based on tensor-enhanced graph similarity according to claim 1, characterized in that, The graph interaction modeling module achieves end-to-end joint training through an embedded optimization layer, including: By introducing auxiliary variable Z and Lagrange multipliers G, an augmented Lagrange function is constructed. The optimization problem is expanded into a network structure with a fixed number of layers using the ADMM alternating update method, and an embedded optimization sub-network is constructed.

3. The drug resistance prediction system based on tensor-enhanced graph similarity according to claim 1, characterized in that, The similarity matrix learning module includes: A multi-scale cross-node similarity coding layer is used to perform multi-scale convolution on the similarity matrix using cross-shape filters of various sizes to obtain feature maps at different scales, and then fuses the feature maps using a weighted average method to obtain a fused feature map. A node similarity learning layer is used to aggregate cross-node information by employing one-dimensional convolution along the node dimension in the fused feature map, generating node-level similarity embeddings; A fully connected layer is used to embed the node similarity into a final graph similarity score.

4. The drug resistance prediction system based on tensor-enhanced graph similarity according to claim 1, characterized in that, The system also includes: The semantic alignment module is used to perform feature transformation and projection on the similarity matrices of multiple semantic perspectives output by the graph interaction modeling module, so as to unify the similarity expression of the similarity matrices.

5. A method for predicting drug resistance based on tensor-enhanced graph similarity, characterized in that, Includes the following steps: To obtain compound and three-dimensional structural data of antiviral drugs; The compound and three-dimensional structural data of the antiviral drug are preprocessed to construct a graphical structure; A node embedding matrix is ​​obtained by learning the initial node features in the graph structure through graph convolution and graph self-attention mechanisms; an attention mechanism is introduced to weight active groups during the node embedding learning stage; specifically, a local embedding matrix of nodes is generated by aggregating the features of nodes and their neighboring nodes; the output local embedding matrix is ​​represented as follows: ； in, This represents the adjacency matrix with added self-loops, where A is the original adjacency matrix and IN is the N-dimensional identity matrix. For symmetric normalization, For degree matrix, ; This is the feature matrix of the nodes in the (l-1)th layer. For trainable weight matrix, H is a nonlinear activation function; H(l-1) is the residual term; The dependencies between nodes are captured using a multi-head self-attention mechanism, and the node embedding matrix is ​​obtained by combining the local embedding matrix; the node embedding matrix is ​​represented as: ； in, , , Let represent the query, key, and value matrices of the m-th attention head, respectively, and dk be the scaling factor for the dimension of the key vector. The learnable structural correlation coefficient. For the degree matrix bias term based on graph topology; Based on the t-product, a multi-view tensor is constructed from the node embedding matrix, and the interaction between nodes in the graph is performed through tensor quantum space learning to obtain a similarity matrix from multiple semantic perspectives; specifically, the node features corresponding to the nodes in the graph are tensored into tensors based on the t-product operation. The tensor is constrained based on the objective function to obtain a self-expressive tensor, and a similarity matrix with multiple semantic perspectives is obtained by accumulating the self-expressive tensors of each dimension. Predict the similarity score of the graph structure based on the similarity matrix; The drug resistance of the antiviral drug is output based on the similarity score.

Images