Drug resistance prediction system and method based on tensor enhancement graph similarity

Abstract

The invention discloses a drug resistance prediction system and method based on tensor enhancement graph similarity, and belongs to the technical field of computer-aided drug detection.The method comprises the steps that an acquisition module obtains drug compounds and three-dimensional structure data, and a preprocessing module constructs a graph structure; the node embedding learning module learns node features through graph convolution and a graph self-attention mechanism, the graph interaction modeling module performs inter-graph node interaction based on t-product, the similarity matrix learning module predicts a graph structure similarity score, and the result output module outputs a drug resistance result according to the score. According to the method, through technologies such as multi-view tensor modeling and node embedding learning, drug structure features are comprehensively captured, and the prediction accuracy is improved; the calculation efficiency is improved by means of end-to-end joint training, multi-scale convolution and the like; the generalization ability of the model is enhanced through preprocessing starting from original data and graph structure construction; the similarity matrix of multiple semantic perspectives provides rich information for drug resistance evaluation.

Description

Technical Field

[0001] The present invention relates to the technical field of computer-aided drug detection, and more particularly to an intelligent prediction system and method for antiviral drug resistance based on tensor-enhanced graph similarity calculation. Background Art

[0002] In recent years, graph neural networks (GNNs) have been gradually introduced into tasks such as molecular similarity calculation and molecular activity prediction. They have overcome the limitations of traditional molecular fingerprinting methods in terms of structural representation and prediction accuracy, becoming a key research direction in drug discovery. Compared to SMILES-based or molecular fingerprint-based descriptions, GNNs can more naturally model the topological relationships and complex atomic interactions 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 for drug resistance prediction of highly mutating viruses like HIV, existing methods still have the following key issues, which seriously restrict their practicality and generalization capabilities: First, existing methods fail to fully model the three-dimensional complementary nature of the HIV protease target. In actual molecule-target interactions, the spatial fit between the molecular conformation and the target active site plays a crucial role in determining drug efficacy. Traditional two-dimensional graph modeling methods often overlook this three-dimensional structural docking relationship. As a result, even with similar molecular topologies, differences in conformation can still lead to significant differences in drug efficacy, thus affecting the model's accurate assessment of drug activity and resistance.

[0004] Secondly, existing graph neural networks (GNNs) lack the ability to model molecular functional groups, failing to effectively capture key pharmacophore features such as hydrogen bond donors and acceptors. Functional groups play a central role in the binding of molecules to viral targets, with non-covalent interactions such as hydrogen bonds, in particular, crucial for the stability of the molecule-target complex. However, current GNN architectures generally lack explicit modeling of functional group semantics, making it difficult to accurately capture the molecular segments that drive drug efficacy, impacting the model's predictive power and reliability.

[0005] Third, existing methods also have significant shortcomings in target-specific modeling. General graph similarity metrics fail to adequately capture the dynamic binding pocket characteristics of targets like HIV protease. When applied across different targets (e.g., HIV-PR and HIV-RT), the early enrichment capacity (EF1%) of existing models decreases significantly, limiting their practicality and broad-spectrum application in multi-target synergistic antiviral drug screening.

[0006] In addition, traditional methods generally lack a mechanism to focus on active substructures. They usually treat all substructures in a molecule equally, ignoring the retention of key structural motifs such as protease_hinge. As a result, some molecules containing invalid or redundant fragments (such as flexible long chains) are overestimated in their similarity, increasing the probability of screening false candidates and reducing the efficiency and accuracy of drug screening.

[0007] In practical applications, there's still an imbalance between efficiency and accuracy. While molecular docking-based methods offer high predictive accuracy, they're computationally expensive, and docking analysis of a single molecule is time-consuming, making them difficult to support high-throughput screening of large-scale molecular libraries. Traditional methods, such as molecular fingerprinting, while computationally efficient, lack the ability to fully capture the three-dimensional structure and functional group characteristics of complex molecules, resulting in insufficient prediction accuracy and a struggle to meet the demands of predicting drug resistance in complex viruses.

[0008] In summary, how to provide an intelligent prediction method and system for antiviral drug resistance that can integrate strong structural feature modeling capabilities, has target adaptability, and supports efficient inter-graph similarity measurement and multi-scale feature extraction has become a technical problem that needs to be urgently solved in this field. Summary of the Invention

[0009] In view of this, the present invention provides a drug resistance prediction system and method based on tensor-enhanced graph similarity. Through four stages of node embedding learning, graph interaction modeling, similarity matrix alignment and similarity matrix learning, it achieves accurate characterization of three-dimensional pharmacodynamic characteristics, targeted focusing of active substructures, enhanced adaptability of drug-resistant mutations, efficient screening of target specificity, and rapid retrieval of large-scale libraries.

[0010] In order to achieve the above object, the present invention adopts the following technical solutions: In one aspect, the present invention provides a drug resistance prediction system based on tensor-enhanced graph similarity, comprising: Acquisition module, used to obtain compound and three-dimensional structure data of antiviral drugs; A preprocessing module, used to preprocess the compound and three-dimensional structure data of the antiviral drug to generate a graph structure; A node embedding learning module is used to learn the initial node features in the graph structure through graph convolution and graph self-attention mechanism to obtain a node embedding matrix; A graph interaction modeling module is used to construct a multi-perspective tensor for the node embedding matrix based on the t-product, and perform node interaction between graphs through tensor subspace learning to obtain similarity matrices from multiple semantic perspectives; A similarity matrix learning module, configured to predict a similarity score of the graph structure based on the similarity matrix; A result output module is used to output the drug resistance of the antiviral drug based on the similarity score.

[0011] Preferably, the node embedding learning module includes: The residual graph convolution unit is used to generate the local embedding matrix of a node by aggregating the features of the node and its adjacent nodes; The graph attention unit is used to capture the dependencies between nodes through a multi-head self-attention mechanism and obtain a node embedding matrix by combining the local embedding matrix.

[0012] Preferably, the graph interaction modeling module includes: A node embedding construction unit, configured to convert the node features corresponding to the graph pair data into a tensor based on a t-product operation; The tensor subspace learning unit is used to constrain the tensor based on the objective function to obtain a self-expression tensor, and to obtain similarity matrices of multiple semantic perspectives by accumulating the self-expression tensors of each dimension.

[0013] Preferably, the graph interaction modeling module implements end-to-end joint training through an embedded optimization layer, including: Introduce auxiliary variables Z and Lagrange multipliers G to construct augmented Lagrangian function; The optimization problem is expanded into a network structure with a fixed number of layers through the ADMM alternating update method, and an embedded optimization sub-network is constructed.

[0014] Preferably, the similarity matrix learning module includes: A multi-scale cross-node similarity encoding layer is used to perform multi-scale convolution on the similarity matrix using cross-shaped filters of various sizes to obtain feature maps at different scales, and fuse the feature maps by weighted averaging to obtain a fused feature map; A node similarity learning layer is used to aggregate cross-node information using one-dimensional convolution along the node dimension in the fused feature graph to generate node-level similarity embedding; A fully connected layer is used to map the node similarity embedding into a final graph similarity score.

[0015] Preferably, the system further comprises: The semantic alignment module is used to perform feature conversion and projection on the similarity matrices of multiple semantic perspectives output by the graph interaction modeling module, and unify the similarity expressions of the similarity matrices.

[0016] In another aspect, the present invention provides a method for predicting drug resistance based on tensor enhancement graph similarity, comprising the following steps: Obtain compound and three-dimensional structure data of antiviral drugs; Preprocessing the compound and three-dimensional structure data of the antiviral drug to generate a graph structure; The initial node features in the graph structure are learned through graph convolution and graph self-attention mechanism to obtain a node embedding matrix; Constructing a multi-perspective tensor for the node embedding matrix based on t-product, and performing node interaction between graphs through tensor subspace learning to obtain similarity matrices from multiple semantic perspectives; Predicting a similarity score of the graph structure according to the similarity matrix; The drug resistance of the antiviral drug is output based on the similarity score.

[0017] It can be seen from the above technical solutions that, compared with the prior art, the present invention discloses a drug resistance prediction system and method based on tensor-enhanced graph similarity. First, through multi-perspective tensor modeling, node embedding learning, and semantic alignment, the structural characteristics of drugs are comprehensively captured from multiple perspectives to improve prediction accuracy. Secondly, end-to-end joint training and multi-scale cross-node similarity encoding and other means have improved computational efficiency and accelerated model training and prediction speed. In addition, the combination of preprocessing and graph structure construction based on raw data, as well as graph convolution and graph self-attention mechanisms, enhances the model's generalization ability for different types of drugs, enabling it to maintain good prediction performance in different data sets and application scenarios. At the same time, the similarity matrix of multiple semantic perspectives provides richer information for drug resistance assessment, which helps to deeply understand the mechanism of drug resistance. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0019] Figure 1 This is a schematic diagram of the overall structure provided by the present invention.

[0020] Figure 2 This is a detailed structural diagram provided by the present invention.

[0021] Figure 3 A structural diagram of another embodiment of the present invention is provided.

[0022] Figure 4 A schematic diagram of the process provided by the present invention. DETAILED DESCRIPTION

[0023] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0024] The embodiment of the present invention discloses one aspect, which provides a drug resistance prediction system based on tensor enhancement graph similarity, such as Figure 1-2 Shown, including: The acquisition module is used to obtain compound and three-dimensional structure data of antiviral drugs.

[0025] The preprocessing module is used to preprocess antiviral drug compound and 3D structure data to generate graph structures. The chemical molecular graphs of the acquired antiviral drug compound and 3D structure data can be represented in the SMILES format (Simplified Molecular Input Linear Representation). Data preprocessing requires converting the molecular structure into a graph format, with nodes representing atoms and edges representing chemical bonds. Chemical computing libraries such as RDKit or Deep Chem are used to read and process these molecular graphs.

[0026] The node embedding learning module uses graph convolution and graph self-attention to learn the initial node features in the graph structure and generate a node embedding matrix. The pharmacophore attention mechanism is introduced during the node embedding learning phase, increasing the model's embedding weight for key active groups (such as the hydroxyethylene isostere of HIV protease inhibitors) by 3-5 times, significantly reducing interference from invalid fragments and lowering the false positive rate.

[0027] The graph interaction modeling module constructs a multi-perspective tensor based on the node embedding matrix using the t-product. It then uses tensor subspace learning to learn node interactions between graphs, yielding similarity matrices for multiple semantic perspectives. Specifically, based on the acquired node embedding data, a multi-perspective tensor is constructed based on the t-product, and node interactions between graphs are learned using tensor subspace learning. After applying low-rank constraints and graph regularization, not only is information within a single graph fused, but multiple semantic perspective similarity matrices are also obtained, characterizing cross-graph interactions between two graphs.

[0028] Similarity matrix learning module, used to predict the similarity score of the structure based on the similarity matrix; The result output module outputs antiviral drug resistance predictions based on similarity scores. Prediction results include recommendations for optimizing anti-HIV activity: When similarity scores exceed 0.7 and the protein contains the "protease_hinge" substructure, drug-resistance mutation testing is recommended. Molecules with similarities between 0.5 and 0.7 are labeled "backbone transition candidates," and a library of side-chain substitution proposals is provided. The output is visualized, including molecular alignment overlays and pharmacophore matching heatmaps.

[0029] Specifically, the node embedding learning module includes: The residual graph convolution unit is used to generate the local embedding matrix of the node by aggregating the features of the node and its adjacent nodes. To avoid the problem of over-smoothing, the residual connection is introduced. The output local embedding matrix is expressed as: ; in, Represents the adjacency matrix with self-loop added (A is the original adjacency matrix, I N is an N-dimensional identity matrix), and the node’s own characteristics are retained by introducing self-loops; Symmetric normalization is implemented ( is the degree matrix, ), in order to eliminate the bias of node degree difference on feature propagation; is the node feature matrix of the l-1 layer, through the trainable weight matrix Complete the feature space projection and pass it through the nonlinear activation function Introducing model expression capabilities; the final residual term H (l-1) By retaining the original feature information through cross-layer connections, the vanishing gradient problem is effectively alleviated. This allows for the synergy of neighborhood feature aggregation, nonlinear transformation, and residual learning to achieve efficient representation learning of graph-structured data.

[0030] The graph attention unit is used to capture the dependencies between nodes through the multi-head self-attention mechanism and combine it with the local embedding matrix to obtain the node embedding matrix. The node embedding matrix is expressed as: ; in, , , Represent the query, key, and value matrices of the mth attention head, respectively. d k is the dimension scaling factor of the key vector (used to stabilize gradient calculation), is the learnable structural correlation coefficient, is the degree matrix bias term based on the graph topology (encoding the connection strength between nodes), the Softmax function normalizes the attention score along the row direction, and the final output is obtained by fusing the node feature similarity (by representation) and graph structure prior information (by rD g This calculation mechanism enhances the adaptability of attention weights to network topology by explicitly incorporating spatial constraints of the graph.

[0031] Furthermore, the graph interaction modeling module includes: The node embedding construction unit is used to convert the node features corresponding to the graph data into tensors based on the t-product operation; for example, the node embedding matrices of graphs A and B are: , , where N A With N B are the number of nodes in graph A and graph B respectively, and d represents the feature dimension. The node features of the two graphs are stretched into tensors: ; In this tensor, the first horizontal slice (frontal slice) corresponds to the information of Figure A, and the second horizontal slice corresponds to the information of Figure B.

[0032] The tensor subspace learning unit is used to constrain the tensor based on the objective function to obtain the self-expression tensor, and to obtain the similarity matrix of multiple semantic perspectives by accumulating the self-expression tensors of each dimension. Can be expressed by self-expression tensor Approximate reconstruction, that is: ; in Defined as t-product: ; Expand the tensor C in the third dimension; express The block circulant matrix of ; Reconstruct the product result into the original tensor format.

[0033] The constructed self-expression tensor C contains the data of two horizontal slices, where: It can mainly reflect the interaction within Figure A and between Figure A and Figure B; It reflects the interaction within Figure B and between Figure B and Figure A.

[0034] To ensure that the self-expression tensor C has both global low-rank properties and preserves the local graph structure, the objective function is defined as follows: ; in, represents the Frobenius norm; represents the tensor nuclear norm (TNN), which is used to force C to have a low-rank structure. ,in is the i-th transverse slice data on the third dimension of the three-dimensional tensor after Fourier transform of C; Represents the vth transverse slice data; is the graph Laplacian matrix corresponding to the perspective v, which is constructed based on the similarity between node embeddings; Parameter θ and To control the weight of low rank and graph regularization terms.

[0035] By performing absolute value calculations on each horizontal slice and performing transposed accumulation: ; Thus, a similarity matrix that integrates the interaction information between graph A and graph B is obtained. At the same time, each horizontal slice data is retained as a representation of different semantic perspectives, namely: , forming a similarity matrix under multiple semantic perspectives.

[0036] Furthermore, in order to enable the above optimization problem to be jointly trained with the neural network end-to-end, the present invention implements end-to-end joint training by embedding the optimization layer in the graph interaction modeling module, including: Introduce auxiliary variables Z and Lagrange multipliers G to construct augmented Lagrangian function; The optimization problem is expanded into a network structure with a fixed number of layers through the ADMM alternating update method, and an embedded optimization sub-network is constructed.

[0037] The specific steps are: Introduce auxiliary variables Z (used to replace C to solve the TNN term) and Lagrange multiplier G to construct the augmented Lagrangian function: ; Among them, μ is the penalty parameter, Represents the inner product operation.

[0038] The above optimization problem is expanded into a network structure with a fixed number of layers using the ADMM alternating update method. Each layer t includes the following update steps: In the forward propagation, each Fourier domain frontal slice is updated using the singular value threshold (SVT) method: ; After inverse FFT, Z is obtained t+1 .

[0039] Solve the following equation: ; This problem can be solved by decomposing each frontal slice in the FFT domain to obtain a closed-form or iterative update formula, and ensure that the update result is consistent with the output of the forward pass layer.

[0040] ; By fixedly expanding the above update process by T layers, all parameters of each layer can participate in backpropagation, forming a trainable embedded optimization subnetwork.

[0041] The present invention adopts a dynamic weight adjustment strategy to optimize graph interaction modeling parameters for HIV protease targets, thereby improving the early enrichment rate (EF1%).

[0042] Furthermore, the similarity matrix learning module includes: The multi-scale cross-node similarity encoding layer is used to perform multi-scale convolution on the similarity matrix using cross-shaped filters of various sizes to obtain feature maps at different scales, and fuse the feature maps by weighted averaging to obtain a fused feature map. Specifically, cross-shaped filters of various sizes are designed. For example: Crossover filter of size 3×3 : Only the middle row and middle column are valid; Crossover filter of size 5×5 ; Crossover filter of size 7×7 ; The formula of each filter is as follows (taking 3×3 as an example): ; Among them, w ij It is a filter The learnable weight parameters are located at the center cross position; Perform multi-scale convolution on the similarity matrix separately: ; in, Represents different sizes, and the output feature map shape is , C s is the number of output channels at the corresponding scale.

[0043] The feature maps generated at different scales are fused into a unified representation through weighted averaging: ; in are trainable weight parameters.

[0044] The node similarity learning layer is used to fuse feature maps and aggregate cross-node information using one-dimensional convolution along the node dimension to generate node-level similarity embeddings; The fully connected layer is used to embed and map the node similarity into the final graph similarity score. Specifically, the input is the fused multi-scale CSL layer output. ,in Indicates the total number of channels at all scales.

[0045] A 1D convolution along the node dimension (usually row or column direction) is used to aggregate cross-node information.

[0046] For each node, its corresponding row (or column) is considered as a sequence and a 1D convolution filter F is applied NSL : ; in, is the node-level similarity embedding, Denotes the activation function, and b is the bias term. Here, 1D convolution will capture the local aggregation features of each node at different locations (across nodes). , is a one-dimensional convolution filter; k is the width of the convolution kernel (sliding window size), which controls the coverage of the local neighborhood. is the number of input channels, corresponding to multi-scale convolution features The total number of channels, d is the number of output channels (ie the number of filter banks), which determines the final node embedding dimension.

[0047] In another embodiment, in order to make full use of the similarity matrices from different perspectives and ensure their consistency in the common semantic space, as shown in FIG. Figure 3 As shown, the present invention introduces a semantic alignment module to perform feature conversion and projection on the similarity matrices of multiple semantic perspectives output by the graph interaction modeling module, unifying the similarity expression of the similarity matrices. This embodiment integrates mutation sensitivity coefficients during the similarity matrix alignment phase to accurately predict the impact of common drug-resistant mutations such as V82A and I84V on molecular activity. Specifically, it includes: For each view v, set the trainable linear mapping matrix , the original similarity matrix Mapping to a common semantic space: ,in, Represents the similarity matrix after semantic alignment.

[0048] In order to ensure consistency in the alignment results of different perspectives, the following alignment loss is designed: ; Its purpose is to minimize the differences between perspectives and thus establish consistent similarity expressions in a common semantic space.

[0049] For multiple aligned view matrices, a unified similarity matrix is obtained by weighted fusion: ; in is a learnable weight parameter that reflects the importance of each semantic perspective.

[0050] On the other hand, the present invention provides a drug resistance prediction method based on tensor enhancement graph similarity, such as Figure 4 As shown, the following steps are included: Obtain compound and three-dimensional structure data of antiviral drugs; Preprocessing of antiviral drug compounds and three-dimensional structure data to generate graph structures; The node embedding matrix is obtained by learning the initial node features in the graph structure through graph convolution and graph self-attention mechanism; Constructing a multi-perspective tensor for the node embedding matrix based on t-product, and performing node interaction between graphs through tensor subspace learning to obtain similarity matrices from multiple semantic perspectives; Predict the similarity score of the graph structure based on the similarity matrix; Outputs antiviral drug resistance based on similarity scores.

[0051] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the devices disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0052] The above description of the disclosed embodiments is intended to enable one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be readily apparent to one 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 present invention. Therefore, the present invention is not limited to the embodiments shown herein but is intended to conform to 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: Acquisition module, used to obtain compound and three-dimensional structure data of antiviral drugs; A preprocessing module, used to preprocess the compound and three-dimensional structure data of the antiviral drug to generate a graph structure; A node embedding learning module is used to learn the initial node features in the graph structure through graph convolution and graph self-attention mechanism to obtain a node embedding matrix; A graph interaction modeling module is used to construct a multi-perspective tensor for the node embedding matrix based on the t-product, and perform node interaction between graphs through tensor subspace learning to obtain similarity matrices from multiple semantic perspectives; A similarity matrix learning module, configured to predict a similarity score of the graph structure based on the similarity matrix; A result output module is used to output the drug resistance of the antiviral drug based on the similarity score.

2. A drug resistance prediction system based on tensor enhanced graph similarity according to claim 1, characterized in that: The node embedding learning module includes: The residual graph convolution unit is used to generate the local embedding matrix of a node by aggregating the features of the node and its adjacent nodes; The graph attention unit is used to capture the dependencies between nodes through a multi-head self-attention mechanism and obtain a node embedding matrix by combining the local embedding matrix.

3. The drug resistance prediction system based on tensor enhanced graph similarity according to claim 1, characterized in that: The graph interaction modeling module includes: A node embedding construction unit, configured to convert the node features corresponding to the graph pair data into a tensor based on a t-product operation; The tensor subspace learning unit is used to constrain the tensor based on the objective function to obtain a self-expression tensor, and to obtain similarity matrices of multiple semantic perspectives by accumulating the self-expression tensors of each dimension.

4. The drug resistance prediction system based on tensor enhanced graph similarity according to claim 3, characterized in that: The graph interaction modeling module implements end-to-end joint training through an embedded optimization layer, including: Introduce auxiliary variables Z and Lagrange multipliers G to construct augmented Lagrangian function; The optimization problem is expanded into a network structure with a fixed number of layers through the ADMM alternating update method, and an embedded optimization sub-network is constructed.

5. 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 encoding layer is used to perform multi-scale convolution on the similarity matrix using cross-shaped filters of various sizes to obtain feature maps at different scales, and fuse the feature maps by weighted averaging to obtain a fused feature map; A node similarity learning layer is used to aggregate cross-node information using one-dimensional convolution along the node dimension in the fused feature graph to generate node-level similarity embedding; A fully connected layer is used to map the node similarity embedding into a final graph similarity score.

6. The drug resistance prediction system based on tensor enhanced graph similarity according to claim 1, characterized in that: The system further comprises: The semantic alignment module is used to perform feature conversion and projection on the similarity matrices of multiple semantic perspectives output by the graph interaction modeling module, and unify the similarity expressions of the similarity matrices.

7. A drug resistance prediction method based on tensor enhancement graph similarity, characterized in that: The following steps are involved: Obtain compound and three-dimensional structure data of antiviral drugs; Preprocessing the compound and three-dimensional structure data of the antiviral drug to generate a graph structure; The initial node features in the graph structure are learned through graph convolution and graph self-attention mechanism to obtain a node embedding matrix; Constructing a multi-perspective tensor for the node embedding matrix based on t-product, and performing node interaction between graphs through tensor subspace learning to obtain similarity matrices from multiple semantic perspectives; Predicting a similarity score of the graph structure according to the similarity matrix; The drug resistance of the antiviral drug is output based on the similarity score.

Images