Drug interaction prediction method based on multi-modal molecular characterization

By combining edge-aware GCNII and graph-attention GATConv for joint representation learning, and integrating multimodal feature fusion of Transformer and Mol2Vec/MolT5, the problems of molecular characterization and transsubstructure interaction in drug interaction research are solved, achieving high-precision and interpretable drug interaction prediction, and supporting drug safety assessment and new drug development.

CN121687172APending Publication Date: 2026-03-17JIANGNAN UNIV
View PDF 0 Cites 2 Cited by

Patent Information

Application Number
CN202511840258.7
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2025-12-08
Publication Date
2026-03-17

AI Technical Summary

Technical Problem

Existing methods for studying drug interactions suffer from limitations in molecular characterization, insufficient cross-substructure interactions, difficulties in fusing multimodal features, and insufficient model generalization and interpretability, making it difficult to meet the needs of large-scale, high-throughput drug interaction prediction.

Method used

We employ edge-aware GCNII combined with graph attention GATConv for joint node-edge representation learning, combine Transformer to capture higher-order dependencies, integrate molecular local substructure context information from Mol2Vec and MolT5 to construct intra- and inter-graph interaction networks, design a multimodal DDI predictor, and use SHAP feature selection to compress key features and enhance interpretability.

Benefits of technology

It achieves high-precision prediction of drug interactions, improves the model's generalization ability and interpretability, and can reveal potential biological mechanisms, providing reliable decision support for clinical multidrug combination therapy and new drug development.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN121687172A_ABST
    Figure CN121687172A_ABST
Patent Text Reader

Abstract

The invention belongs to the technical field of artificial intelligence algorithm and bioinformatics crossing, and relates to a drug interaction prediction method based on multi-modal molecular characterization. According to the method, through the edge perception GCNII architecture and the Hop2Token multi-hop coding mechanism, effective modeling of atomic-level and bond-level local environments and a cross-substructure high-order dependency relationship in drug molecules is realized, and the accuracy and robustness of drug interaction prediction are improved. According to the method, Mol2Vec and MolT5 cross-modal molecular characterization is integrated, fusion of molecular overall semantics and substructure grammar semantic association is achieved, and the generalization ability of the model to complex molecules and unknown medicine combinations is remarkably improved. According to the method, the dynamic feature screening algorithm driven by the SHAP value of the artificial intelligence technology is adopted for biological verification, the feature redundancy problem is effectively solved, the molecular biological information analysis processing calculation efficiency and the model transparency are improved, and the traceability of the prediction process is guaranteed.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention belongs to the interdisciplinary field of artificial intelligence algorithms and bioinformatics, and relates to a method for predicting drug interactions based on multimodal molecular characterization. Background Technology

[0002] Drug-drug interaction (DDI), an important topic in pharmacological research, refers to the synergistic or antagonistic effects that occur when two or more drugs act simultaneously in the body. These combined effects can either enhance efficacy or reduce adverse reactions, or weaken drug efficacy or cause unexpected toxicity, posing significant challenges to drug safety evaluation, new drug development, and clinical medication regimen design.

[0003] Current drug interaction (DDI) studies primarily rely on traditional in vivo and in vitro experimental methods, assessing potential risks through changes in blood drug concentrations, pharmacokinetic analysis, and clinical observation. However, these methods suffer from problems such as insufficient sample representativeness, unavoidable detection errors, and high time and cost, making it difficult to meet the demands of modern drug development for large-scale, high-throughput prediction. In recent years, artificial intelligence technologies, especially deep learning methods, have provided new avenues for DDI prediction. Traditional machine learning methods based on drug similarity matrices, network reasoning frameworks integrating multi-source heterogeneous data, knowledge graphs, and graph neural networks (GNNs) have improved prediction accuracy to some extent. However, existing research still has significant limitations: 1. Limitations of molecular characterization: Most methods treat drug molecules as a whole, ignoring the dynamic relationships and atomic-level interactions between substructures, making it difficult to accurately resolve the mechanisms of action of key pharmacophores and substructures.

[0004] 2. Insufficient attention to transsubstructure interactions: Existing models do not pay enough attention to the semantic and topological interactions between groups or transsubstructures, making it difficult to fully capture the complex interaction networks between drug molecules and limiting the precision and interpretability of predictions.

[0005] 3. Challenges in multimodal feature fusion: Drug molecules contain both sequence information (such as SMILES) and graph structure information. Traditional methods suffer from problems such as information redundancy, low dimensionality reduction efficiency, and insufficient model transparency when integrating multimodal features.

[0006] 4. Insufficient model generalization and interpretability: Existing methods still have limited predictive performance and robustness when faced with unknown drugs or limited samples, and lack traceable mechanistic explanations, making it difficult to provide scientific evidence for clinical drug safety.

[0007] Therefore, there is an urgent need to develop a novel intelligent tool that can: on the one hand, fully utilize artificial intelligence technology for comprehensive biovalidation to uncover dynamic correlations and cross-substructure interactions between molecular substructures; on the other hand, enable multimodal feature fusion, redundancy optimization, and efficient modeling in molecular bioinformatics analysis systems; and simultaneously possess strong generalization ability and good interpretability. This tool can not only accurately predict drug interactions but also provide reliable decision support for clinical multidrug combination therapy and new drug development. Summary of the Invention

[0008] The technical solution of the present invention is as follows: This invention proposes a drug interaction prediction method based on multimodal molecular characterization. To achieve high-precision modeling of drug molecule structure and interactions, this invention first utilizes edge-aware GCNII combined with graph attention GATConv, applying artificial intelligence technology, to perform node-edge joint characterization learning on the drug molecule graph. This effectively captures local chemical bond information of biological mechanisms and molecular features of global topological structures. Simultaneously, multi-hop neighborhoods are serialized into tokens, and higher-order dependencies are captured using Transformer to achieve cross-neighborhood structural feature learning. Furthermore, this invention integrates molecular local substructure context information captured by Mol2Vec with cross-modal semantic embedding provided by MolT5, and combines SHAP feature selection for key feature compression and enhanced interpretability. In terms of interaction modeling, this invention constructs intra-graph and inter-graph interaction networks, utilizing Graph-intra and Graph-Inter to enhance the consistency of internal drug structures and the molecular features of inter-drug interactions. To address the task of predicting complex drug-drug interactions, this invention designs a multimodal DDI predictor. The graph representation and interaction features mentioned above are input into the predictor for joint learning and classification. At the same time, based on interpretable signals at the atomic, chemical bond, and multi-hop neighborhood levels, the model can not only predict whether there are interactions between drugs, but also reveal the underlying biological mechanisms.

[0009] The drug interaction prediction method based on multimodal molecular characterization consists of the following steps: Step 1: Dataset Construction To develop statistical prediction models, constructing reliable training and testing datasets is crucial; therefore, a meticulous and rigorous strategy was adopted. The specific steps are as follows: 1.1 Data Collection and Positive / Negative Sample Construction This invention collects two publicly available drug identification and treatment (DDI) benchmark datasets: DrugBank and ChChMiner. The DrugBank dataset, derived from FDA and Health Canada drug labels, contains 1710 drugs and 192,284 interactions, covering 86 interaction types. Since DrugBank only contains positive samples, this invention employs a Bernoulli distribution-based negative sample generation method proposed by Shen et al. to generate an equal number of negative samples to achieve a balance between positive and negative samples. The ChChMiner dataset contains 998 drugs and 33,669 interaction records, including 21,486 positive samples and 12,183 negative samples. It includes 798 known drugs and 199 unknown drugs, and the interaction types are binary classification.

[0010] 1.2 Data partitioning and evaluation tasks To comprehensively evaluate model performance, this invention designed a propagation task and an induction task on two datasets: DrugBank and ChChMiner. In the propagation task, the DrugBank dataset used 3-fold cross-validation, with the training, validation, and test sets divided in a 3:1:1 ratio. The ChChMiner dataset also used a 3:1:1 ratio for the training, validation, and test sets. In the induction task, to simulate a new drug prediction scenario, drugs in both datasets were randomly divided into new drugs (20%) and old drugs (80%). The training set contained only old drugs, while the test set was further divided into S1 (interactions between new drugs) and S2 (interactions between new and old drugs). This partitioning process was repeated three times to reduce the impact of random splitting, thereby ensuring the model's generalization ability and robustness under limited sample and unknown drug conditions.

[0011] Step 2: Construction method of feature extraction module This invention proposes a drug feature extraction method that integrates multimodal information. This method uniformly encodes the molecular graph topological attributes and global semantic information of a drug, thereby obtaining a more expressive multimodal representation of the drug and providing a reliable feature foundation for subsequent drug-drug interaction prediction.

[0012] 2.1 Graph Structure Feature Encoding Module The graph structure encoding module parses the SMILES string of the drug using the RDKit tool and generates a molecular graph structure representation. This molecular graph includes a set of atoms V and a set of edges E. The chemical features of each atom and edge are defined according to the knowledge system of the chemical field and are linearly mapped and nonlinearly activated through a learnable parameter matrix, thereby forming initial atomic feature embeddings and edge feature embeddings. This invention introduces the ReLU activation function to achieve nonlinear enhancement of features, enabling the full expression of the topological structural features of the drug molecule at the atomic level.

[0013] 2.2 Global Molecular Descriptor Construction Module To compensate for the limitations of graph structure features, this invention further introduces a global molecular descriptor generation mechanism. This mechanism integrates two unsupervised molecular characterization techniques: Mol2Vec and MolT5. The Mol2Vec module, analogous to word embeddings in natural language, maps molecular substructures to a continuous vector space, thus forming a global representation reflecting chemical function and structural similarity. The MolT5 module, based on a multi-layer attention mechanism of the Transformer architecture, extracts semantically relevant contextual information from the molecular SMILES sequence, generating deep molecular embeddings with high-dimensional semantic features. The combination of these two mechanisms enhances the ability to model drug semantics and reaction patterns while maintaining the integrity of the chemical structure.

[0014] Step 3: Feature selection and dimensionality reduction based on SHAP Considering the redundancy in computation and generalization of high-dimensional features generated by the MolT5 encoder, this invention proposes a supervised feature selection method based on the SHAP interpretation mechanism. This method first constructs a drug interaction prediction framework using the LightGBM model, then calculates the SHAP contribution value of each feature dimension and sorts them in descending order of their absolute mean. Finally, the top 300 key features are selected for subsequent modeling, achieving feature compression from 1024 dimensions to 300 dimensions. This method effectively improves the discriminativeness and interpretability of features while ensuring the integrity of the model's representation, thereby significantly improving the efficiency and reliability of the drug interaction prediction model.

[0015] Step 4: Methods for constructing drug interaction prediction models This invention proposes a multi-level drug interaction prediction model. This model, based on a graph neural network, achieves hierarchical feature extraction and interaction discrimination from the atomic to the drug level through multi-scale structural modeling and multi-modal feature fusion. Its core innovations include: introducing a graph attention network (GAT) to dynamically enhance node features; designing an edge-aware GCNII convolutional module, combining initial residuals and identity mapping mechanisms to suppress over-smoothing while utilizing edge feature attention to achieve structural differentiation aggregation; serializing multi-hop neighborhood information into token representations through Hop2Token and Transformer mechanisms to capture global dependencies between different topological levels; combining intra-graph and inter-graph feature extraction mechanisms, utilizing a co-attention mechanism to model complex cross-drug interactions; introducing adaptive subgraph pooling (SAGPooling) for structural compression and key node selection; fusing molecular semantic features (Mol2Vec, MolT5) to construct a unified drug representation; and finally achieving accurate prediction of drug-drug interactions through multi-modal fusion and classification modules. The model framework diagram is shown below. Figure 1 As shown, the model mainly includes the following functional modules: 4.1 Graph Attention Neural Network (GAT) This invention first uses an edge-attribute-aware Graph Attention Network (GAT) to update node features. Unlike standard GCNs that uniformly or statically weight neighbors, GAT dynamically calculates attention weights for each neighbor through a self-attention mechanism. Let... , It is a node and nodes Features; It is a linear mapping weight matrix of node features, used to calculate the node. and nodes Attention coefficient between As shown in equation (1): (1) in, It is a learnable attention vector. This indicates vector concatenation. The attention weights are obtained by normalizing the attention coefficients using the activation function with Softmax. As shown in equation (2): (2) In the formula, It is a node The set of adjacent nodes is used to aggregate the features of adjacent nodes based on attention weights to obtain the node. The expression for the cluster update is shown in equation (3): (3) In the formula, ReLU is a non-linear activation function. :node Update representation in the current layer.

[0016] 4.2 Edge-aware GCNII To address the oversmoothing problem in deep GCN structures, this invention designs an improved GCNII architecture. This module simultaneously introduces initial residual connections and an identity mapping strategy, maintaining deep stability while achieving differentiated aggregation within the neighborhood through an edge feature attention mechanism. This design enables the model to assign differential weights to different types of chemical bonds (single bonds, double bonds, aromatic bonds, etc.), improving the discriminative power of local molecular structures. The GCNII's... ℓ Layers can be formalized as shown in equation (4): (4) Among them, GCNII mainly represents smoothness. With the first layer Combine the initial residual join with an added unit mapping. To the Layer weight matrix This approach aims to suppress excessive smoothing in deep networks while simultaneously improving the stability and trainability of parameter updates.

[0017] To address the importance of bond features in message passing during molecular graph modeling, this invention proposes an Edge-Aware GCNII convolutional module. This module retains the initial residual and identity mapping mechanisms from GCNII to suppress oversmoothing, while introducing learnable attention weights based on bond features. This enables the neighborhood aggregation process to distinguish between different types of bonds. Specifically, the structural and chemical features of each bond feature in the graph are mapped linearly to obtain scalar attention coefficients, which participate in standard normalized neighborhood aggregation. Subsequently, these coefficients are combined with intra-layer coefficients... , The inter-layer residuals and initial features are fused, and the node representation of the next layer is finally output as a nonlinear transformation. The initial feature matrix of the node is as follows: , where N is the number of nodes and d is the feature dimension of the nodes. The node features of the layer are the input features. Each edge The edge features are represented as ,in Let be the edge feature dimension, and let the edge weight be denoted as . Residual coefficient Identity mapping coefficients .

[0018] In this invention, the attention weight of each edge is obtained as a scalar by linearly mapping the edge features. As shown in equation (5): (5) in, For learnable weight vectors, Represents a node The set of neighboring nodes, including self-linking nodes. Attention coefficient. With edge weight Together they are used for neighborhood set. To ensure the stability of information aggregation and to consider self-loops, the adjacency matrix... Add the identity matrix Then, symmetric normalization is performed to obtain the edge-aware normalized adjacency matrix. As shown in equations (6), (7), and (8). Where D is the degree matrix, and the nodes... The neighborhood aggregation representation at layer l is shown in equation (9): (6) (7) (8) (9) Through residual coefficient The original node features and aggregated features are weighted and fused to obtain intermediate features. As shown in equation (10). Then, after applying the identity mapping coefficients... By fusing the intermediate features with the initial features, we obtain the first... Layer output node features As shown in equation (11): (10) (11) Finally, the output is linearly mapped to the target dimension, as shown in equation (12): (12) in, For learnable linear mapping matrix, For bias vectors, Output dimensions for the target.

[0019] In this way, the Edge-Aware GCNII module of the present invention can alleviate the over-smoothing problem in deep networks, and at the same time make full use of edge feature information to achieve differentiated neighborhood aggregation. While preserving the stability of deep structures, it can assign differentiated weights to different types of chemical bonds (such as single bonds, double bonds, and aromatic bonds), thereby enhancing the sensitivity to differences in local molecular structures.

[0020] 4.3 Hop2Token & Transformer To capture high-order structural relationships across neighborhoods, this invention employs the Hop2Token serialization mechanism, transforming multi-hop neighborhood information (k-hop) into equivalent token sequences, which are then input into a Transformer encoder for global dependency modeling. This mechanism can reconstruct multi-level semantic relationships of drug molecule structures in topological space, achieving cross-level learning from local bonding patterns to global structural dependencies, significantly improving the model's sensitivity to complex molecular topologies. The core function of this module is to intelligently convert graph structure data into sequence information; it receives updated node features. As input, through the adjacency matrix of the graph The node features are iteratively aggregated, with each aggregation representing an outward expansion by one hop. The K-hop neighborhood environment of each node is transformed into a sequence, and the powerful sequence processing capabilities of Transformer are used to model the complex relationships between neighborhoods at different distance scales.

[0021] A 0-hop token represents the unique characteristics of a node, namely... k-hop Token represents the first... k The neighbor characteristics are obtained through adjacency matrix multiplication, i.e. , This process is repeated iteratively. (i.e., K times); ultimately, the module aggregates the features of all hop counts. Stack them up to generate a multi-scale feature sequence for each node in the graph: (13) This feature sequence is then fed into the Transformer encoder to further model the dependencies between hop counts. The core of the Transformer consists of Multi-Head Self-Attention (MHSA) and a Position-wise Feed-Forward Network (FFN), designed to capture global contextual information between different hop tokens within the sequence, thereby outputting more expressive context-enhanced representations. In the MHSA module, the self-attention mechanism allows the model to dynamically weigh the importance of other tokens in the sequence when processing a particular token. For this purpose, the input sequence... First it is mapped to the query ,key Sum The three subspaces are then used to calculate the attention weights by scaling the dot product attention; the calculation form is shown in Equation (14): (14) in, , , , Under the multi-head mechanism, features are divided into multiple subspaces for parallel computation, and finally concatenated to form the output, enabling each token to perform weighted aggregation on all hop tokens based on the global context; subsequently, the updated token representation is obtained through LayerNorm and MHSA, as shown in Equation (15). The feedforward network independently performs two-layer linear transformations on each token and adds an activation function to improve the nonlinear modeling capability of the model, and the output sequence is shown in Equation (16).

[0022] (15) (16) To obtain the final vector representation of the nodes, the Transformer encoder ultimately outputs a sequence. After an attention pooling layer, the token sequences representing neighborhoods with different hop counts are aggregated into a single node representation rich in multi-scale structural information. As shown in equations (17) and (18), where, These are trainable parameters, and the contribution of each token is determined by the attention. Decide.

[0023] (17) (18) 4.4 Intra- and Inter-graph Feature Extraction To fully extract the rich features of the drug graph structure, this invention employs a graph attention mechanism, aiming to simultaneously capture the complex interactions between the internal structure of the drug (intra-view) and between different drugs (inter-graph). Specifically, regarding intra-view interactions, we allow atoms (nodes) in the drug graph to send attention information to each other along their connected chemical bonds (edges). Each atom performs information aggregation and feature updates based on the information sent by its neighbors and its own characteristics.

[0024] (19) In equation (19), express and This represents the attention weights between nodes. and Representative node and eigenvectors, and This represents the learnable weight matrix used to calculate the attention score.

[0025] (20) In equation (20) Indicates that the drug is in Node representation after layer GAT. Represents a node The set of neighboring nodes, It is a normalization constant; for feature extraction of inter-graph structures, the core idea of ​​intra-graph feature extraction is also followed.

[0026] (twenty one) (twenty two) As shown in equations (21) and (22), the inter-graph adjacency matrix is ​​represented as a bipartite graph based on the two drugs. Through this construction method, we are able to effectively capture interactions across drug boundaries using a co-attention mechanism. This represents the nodes in the bipartite graph G. The neighbor set of another drug, that is, all atoms in another drug. In other words, allowing drugs... Atoms in another drug Atoms in the process send or receive attention information so that all nodes of another drug can be used to update node information in another drug.

[0027] 4.5 Adaptive Subgraph Pooling (SAGPooling) Structure Information Compression Module To overcome the limitations of existing graph pooling techniques in balancing structure preservation and feature representation, this invention employs an attention-based adaptive subgraph pooling (SAGPooling) structural information compression method. This method learns node importance scores to dynamically filter and adaptively aggregate key structural information, thereby effectively compressing the graph structure size. Its core lies in introducing a two-layer attention mechanism: first, node importance scores are calculated based on node features and neighborhood information; then, high-value nodes are selected based on these scores to construct a subgraph, and the retained nodes are weighted and aggregated to generate the pooled structural representation. This mechanism significantly reduces computational complexity while preserving global semantics and local topological features, improving model scalability and generalization ability.

[0028] The SAGPooling module of this invention can achieve efficient structure compression and information preservation in tasks such as graph classification, node recognition, and drug molecule structure modeling, providing an interpretable and scalable technical solution for the hierarchical representation of complex graph data. The principle is shown in equations (19) and (20): (twenty three) (24) Among them, It is an adjacency matrix. Are all nodes in Embedding in layers, It is the first Layer trainable matrix, It is the input dimension of node-level features.

[0029] 4.6 Multimodal Feature Fusion and Classification Prediction Module This invention aggregates the extracted graph structure features into a graph-level representation using SAGPooling, and then fuses it with semantic features (such as Mol2Vec and MolT5 embeddings) for multimodal fusion. Finally, a fully connected layer and a classifier are used to predict drug-drug interactions. In each block, the final drug graph representation consists of three parts, as shown in equation (25): (25) in, From node pooling, From edge feature pooling, The fusion path from GCNII and Transformer; subsequently, as shown in Equation (26), this figure represents the global molecular features. The components are fused and normalized to obtain the final drug representation of the Block: (26) The results of all blocks are stacked into a three-dimensional tensor: ,inB for , L For the number of blocks, D As the latent space dimension, this tensor is input into the Co-Attention module to perform cross-graph attention interactions between the two drugs, resulting in a cross-drug. Finally, the RESCAL Knowledge Graph Embedding (KGE) module combines the relation vectors... Drug pair interactions are predicted by scoring in a bilinear form, as shown in equation (27): (27) To enhance the model's ability to distinguish between positive and negative samples in drug pair interaction predictions and maintain gradient stability during training, this invention proposes an Adaptive Margin Sigmoid Loss (AMSL) function. This loss function, based on the traditional Sigmoid loss, introduces a temperature adjustment coefficient, a boundary offset term, and a positive / negative sample weight balancing mechanism, thereby enhancing the model's generalization performance in complex drug pair tasks. To further improve the dynamic stability of training, this invention introduces an Adaptive Temperature Scaling mechanism before loss calculation to control the gradient distribution and model convergence rate. The positive sample score is set as follows: Negative sample scoring The temperature coefficient is dynamically adjusted based on the standard deviation of the sample scores. As shown in equation (28). The negative samples are adaptive weights, as shown in equation (29).

[0030] (28) (29) in, Based on the temperature coefficient, For adaptive adjustment coefficients, The standard deviation of the sample scores is represented; the positive and negative sample loss terms are defined as follows: (30) In equation (30), the boundary offset is This is used to increase the discrimination interval between positive and negative samples. for function, Indicates to this The average of the samples is taken, and the positive and negative sample weighting coefficients are used to adjust the contribution of different samples to the loss. The balance coefficient is... The loss function is used to adjust the loss weights and is defined as follows: (31); Through the above mechanism, the loss function of this invention can maintain gradient smoothness and avoid oscillations in the early stage of training, while increasing the separation between positive and negative samples through boundary offset m in the later stage of training, thereby improving the model's convergence speed and classification robustness. This loss function can also effectively prevent gradient oversaturation and numerical instability, significantly enhancing the model's generalization ability in multimodal feature fusion tasks.

[0031] Step 5: Model Training and Validation Testing To verify the effectiveness of the proposed MEGH-DDI model in drug-drug interaction prediction tasks, five core evaluation metrics were selected for comprehensive assessment: accuracy (ACC), area under the receiver operating characteristic curve (AUC), average precision (AP), recall, and F1 score. A systematic comparison with several mainstream baseline models (such as SSI-DDI, GMPNN-CS, DGNN-DDI, DSN-DDI, MIRACLE, and PEB-DDI) was conducted to verify the advantages of the proposed model in multidimensional performance. The experimental design included a transductive task and an inductive task, used to evaluate the model's prediction and generalization abilities in cases of known drug pairs and new drug pairs, respectively.

[0032] Step 6: Interpretable Analysis The drug interaction prediction system proposed in this invention has good interpretability. By introducing a side-level attention analysis mechanism, a multi-hop topological attention mechanism, and an atomic-level feature contribution evaluation method into the model structure, it achieves multi-level interpretation of the drug interaction prediction process. The system includes a data input module, a graph neural network prediction module, and an interpretability analysis module, wherein the interpretability analysis module consists of a side-level attention submodule, a multi-hop topological attention submodule, and an atomic-level contribution analysis submodule.

[0033] The atomic-level contribution analysis submodule calculates the comprehensive contribution score of each atom in the prediction process through attention pooling and weighted aggregation methods, identifies key atoms and their neighboring functional segments, and displays the model's areas of interest through visualization technology, thereby achieving visualization and traceability of the model's decision-making process.

[0034] The edge-level attention submodule is used to calculate the attention weights of each chemical bond in the molecular structure and map high-weight edges to key pharmacophores and potential metabolic sites, thereby revealing the structure-function relationships on which the model predictions depend and achieving consistent structural interpretability. The multi-hop topology attention submodule generates multi-hop (k-hop) attention weights based on the synergistic effect of Hop2TokenEncoder and TransformerEncoderReadout, in order to quantify the degree to which the model depends on information at different topological levels when predicting drug interactions, thereby revealing potential action pathways between drugs across targets, metabolisms, and pharmacodynamics. Through the collaborative design of the above modules, this invention can achieve multi-dimensional interpretability of drug interaction prediction, intuitively display the decision basis of the model, and reveal the source and biological mechanism of the prediction results, thereby improving the transparency, scientific credibility and application verifiability of the model output, and providing reliable technical support for drug mechanism research, new drug development and clinical drug risk assessment.

[0035] The beneficial effects of this invention are: 1. This invention achieves effective modeling of atomic and bond-level local environments and higher-order dependencies across substructures in drug molecules through the edge-aware GCNII architecture and the Hop2Token multi-hop encoding mechanism, thereby improving the accuracy and robustness of drug interaction prediction.

[0036] 2. This invention integrates Mol2Vec and MolT5 cross-modal molecular characterization, realizing the fusion of molecular overall semantics and substructure syntactic semantic association, which significantly improves the model's generalization ability for complex molecules and unknown drug combinations.

[0037] 3. This invention employs a dynamic feature selection algorithm driven by SHAP values, which effectively solves the feature redundancy problem, improves computational efficiency and model transparency, and ensures the traceability of the prediction process.

[0038] 4. This invention constructs a multi-level interpretability analysis module, including edge-level attention, cross-jump topological attention, and atomic-level contribution analysis, which can locate key atoms, pharmacophores, and high-risk chemical bonds, reveal hidden mechanisms of action across targets and pathways, and improve the scientific credibility of prediction results.

[0039] 5. The multi-granularity interpretable mechanism provided by this invention can intuitively visualize the basis for model decision-making, providing a reliable reference for drug mechanism research, DDI risk assessment, and clinical multi-drug combination therapy. Attached Figure Description

[0040] Figure 1 This is a flowchart illustrating the overall technical structure of the present invention.

[0041] Figure 2This invention provides an ablation experiment to systematically evaluate the effectiveness of each module in the proposed model.

[0042] Figure 3 It is a highly valuable atomic visualization result.

[0043] Figure 4 This relates to the impact of different masking strategies on the prediction probability of DDI.

[0044] Figure 5 DDI Key Chemical Bond Attention Visualization

[0045] Figure 6 Dependence distribution of four typical drugs on different jump numbers (k=0 to k=3). Detailed Implementation

[0046] The specific implementation of the present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0047] This invention proposes a drug interaction prediction method based on multimodal deep learning. The method first represents the drug molecule as a graph, constructing a molecular graph with atoms as nodes and chemical bonds as edges. Then, an edge-aware graph neural network (Edge-Aware GCNII) is used to encode the molecular graph, generating feature representations of nodes and edges for subsequent interaction prediction. At the molecular semantic representation level, this invention employs a dual-channel deep learning strategy. A Mol2Vec-based distributed substructure representation module learns substructure co-occurrence patterns from a large-scale molecular corpus, forming a holistic semantic-level representation. A MolT5-based chemical language model module parses the drug's SMILES sequence, extracting chemical grammar and semantic association information, and fusing the two to form a cross-modal embedding space, achieving joint representation of the overall molecular semantics and local substructure features. To reduce feature redundancy and improve model transparency, this invention designs a SHAP-value-driven dynamic feature selection algorithm to intelligently reduce the dimensionality of high-dimensional features while preserving information integrity and improving computational efficiency.

[0048] At the fine-grained structure modeling level, this invention proposes an edge-aware GCNII architecture, introducing an edge-feature-based attention mechanism to distinguish the contributions of different chemical bond types and their local environments to the prediction results. The Hop2Token module transforms multi-hop neighborhood structures in the molecule into token sequences and, combined with the multi-head self-attention mechanism of the Transformer encoder, models high-order dependencies across substructures, thereby obtaining more expressive node representations. In the model training and prediction stages, this invention employs a multi-layer attention mechanism to achieve interpretability analysis. The edge-level attention submodule calculates the attention weights of each chemical bond, mapping high-weight chemical bonds to key pharmacophores and potential metabolic sites, achieving a structure-function consistent interpretation. The multi-hop topological attention submodule generates attention weights with different hop counts to reveal indirect action pathways between drugs across targets, metabolisms, and pharmacodynamics. The atom-level contribution analysis submodule calculates the comprehensive contribution score of each atom through pooling and weighted averaging, identifies key atoms and their neighboring functional fragments, and visualizes the model's focus areas, achieving transparency and traceability in the prediction process. Experimental results show that the present invention achieves a classification accuracy of 98.36% on the DrugBank dataset's propagation task, with both AUC and AP exceeding 99.7%, and an F1 score of 99.05% on the ChChMiner dataset, all significantly outperforming existing baseline methods. In the inductive prediction task, the present invention maintains high prediction performance for unknown drug pairs, validating its superior generalization ability and robustness.

[0049] In summary, the MEGH-DDI method of this invention achieves high accuracy, interpretability, and mechanism consistency in drug interaction prediction through multimodal semantic parsing, edge-aware GCNII, Hop2Token high-order dependency modeling, and multi-level interpretability design, providing an effective tool for drug safety assessment and new drug development.

[0050] Example 1: Dataset Construction and Preprocessing In this invention, a dataset for drug-drug interaction (DDI) prediction is constructed using the DrugBank and ChChMiner databases. The DrugBank dataset is divided into training, validation, and test sets in a 3:1:1 ratio; the ChChMiner dataset is also divided in a 3:1:1 ratio. The original data undergoes deduplication and normalization to ensure the uniqueness and validity of the samples. Subsequently, a dual-view representation method using molecular graphs and molecular fingerprints is employed to extract structural and semantic features of drug molecules, providing high-quality input for subsequent model training and evaluation.

[0051] Example 2: Model Training and Performance Evaluation To verify the effectiveness of the model in this invention, transduction task experiments were conducted on the DrugBank and ChChMiner datasets, evaluating the model's ability to predict interactions between known drugs. The DrugBank dataset used three-fold cross-validation (3:1:1 ratio), and the ChChMiner dataset also used a 3:1:1 split. Experimental results are shown in Tables 1 and 2. On the DrugBank dataset, the traditional methods MHCADDI and GAT-DDI achieved accuracies of 83.80% and 89.81%, respectively, while the accuracy of the model in this invention reached 98.36%, representing improvements of 14.56 and 8.55 percentage points compared to existing methods. AUC and AP reached 99.82% and 99.79%, respectively, demonstrating superior feature representation and relationship modeling capabilities. On the ChChMiner dataset, the model in this invention achieved over 99% in AUC, AP, and F1, with an F1 score of 99.05%, an improvement of 2.74 percentage points compared to the current best model, PEB-DDI (96.31%), demonstrating strong robustness and stability.

[0052] Table 1 Experimental results of the Drugbank dataset on the transduction task. Table 2 Experimental results of the ChChMiner dataset on transduction tasks. Example 3: Inductive Task Experiment The induction task aims to evaluate the model's ability to predict interactions between drugs (new drugs) not seen during training, or between new drugs and known drugs, thus better reflecting the model's generalization ability. Experimental results are shown in Tables 3 and 4. Overall, our model achieved superior performance on both partitions of the two datasets (S1: new drug-new drug; S2: new drug-old drug), significantly outperforming baseline methods. Experimental results on the induction task on the DrugBank and ChChMiner datasets are shown in Tables 3 and 4. Unlike the transduction task, the induction task requires the model to predict drugs or drug pairs not seen during training, thus better reflecting the model's generalization ability. Overall, the model achieved superior performance on both partitions of the two datasets (S1 and S2), significantly outperforming baseline methods.

[0053] On the DrugBank dataset, the model of this invention achieves ACC 76.48%, AUC 84.54%, and AP 84.93% in the Inductive-S1 split; and ACC 83.66%, AUC 92.25%, AP 91.79%, and F1 83.35% in the Inductive-S2 split, both significantly outperforming existing models such as PEB-DDI. On the ChChMiner dataset, the model of this invention achieves AUC 86.18%, AP 92.63%, and F1 87.22% in Inductive-S1; and further improves to AUC 95.36%, AP 97.02%, and F1 92.89% in Inductive-S2, representing improvements of 1.7% to 5.7% over the current best model. In summary, in the inductive task, the model achieves optimal or near-optimal performance in both splits of the two datasets, significantly surpassing existing methods. The results show that the proposed dual-view substructure characterization method is not only applicable to the prediction of interactions between known drugs, but can also be effectively extended to the prediction of new drugs or new drug pairs, demonstrating strong generalization ability and practical potential.

[0054] Table 3 Experimental results of the Drugbank dataset in the inductive task. Table 4 Experimental results of the ChChMiner dataset on the inductive task. Example 4: A Method for Achieving Model Interpretability Based on Multi-Granularity Attention Mechanism This embodiment discloses an interpretability implementation method for a drug interaction prediction model based on a multi-granularity attention mechanism. This method aims to overcome the "black box" limitation of deep learning models in drug interaction prediction. By introducing a multi-granularity attention mechanism, it performs structured analysis of the model's internal reasoning process, thereby revealing the key chemical structures, pharmacological pathways, and topological depth information upon which the model relies in the prediction process, achieving visualization and verifiability of the model's decisions.

[0055] The method comprises the following three aspects: I. Key Atom Identification and Ablation Verification Method Based on Attention Mask This embodiment first identifies key atoms and functional substructures that significantly impact drug-drug interaction (DDI) prediction results based on the model's atomic-level attention weight distribution. Specific visualizations of key atoms are shown below. Figure 3As shown. To verify the scientific rationality of the model's recognition results, a key ablation experiment based on an attention mask was designed. The specific steps included: selecting typical drug combinations (see Table 5) as experimental subjects, including fibrates, statins, antiplatelet drugs, anticoagulants, and psychotropic drugs with CYP inhibitors, etc.; Figure 4 Three experimental conditions were constructed: the complete model (MEGH-DDI); the model with key atoms identified by removing the attention mask; and the control model with randomly removed atoms. The changes in the DDI prediction probability under the three conditions were compared.

[0056] Table 5 Five Typical DDI Cases Experimental results show that the complete model achieves the highest prediction accuracy; removing key atoms significantly reduces prediction performance; while the decrease caused by randomly removing atoms is relatively small. These results demonstrate that key atoms identified by the model's attention mechanism play a crucial role in DDI prediction, providing quantitative and structured evidence to support the model's interpretability. Furthermore, this method can identify key substructures in drug molecules that are highly relevant to pharmacological interactions, providing intuitive scientific evidence for clinical drug combination evaluation and adverse reaction risk prediction.

[0057] II. Structural Interpretability Analysis Methods Based on Chemical Bond Weights This embodiment further analyzes the weight distribution at the edge level of chemical bonds based on the model's edge attention mechanism to reveal the model's response patterns to different chemical structures in DDI prediction. The method introduces an edge attention computation module into the Edge-GCNII framework to achieve a quantitative assessment of the importance of chemical bonds in the molecular diagram. During the prediction process, the model automatically assigns weights to different chemical bonds; edges with high weights typically correspond to key pharmacophores or metabolic sites. For example, ... Figure 5 In the typical pharmacokinetic (PK) interaction between statins (Simvastatin, Atorvastatin) and fibrates (Gemfibrozil), the model assigns higher weights to the carboxylic acid group (–COOH) and ether bond of Gemfibrozil, and to the hydroxyl group (–OH) and lactone ring of statins, reflecting their key role in the competition mechanism of metabolic enzymes.

[0058] Furthermore, in the anticoagulant-antiplatelet scenario of Warfarin combined with Clopidogrel or Aspirin, the model highlights the C=O / C–O bonds of the coumarin backbone in Warfarin and the thiophene ring and ester bond regions in Clopidogrel, corresponding to their pharmacodynamic core and metabolic activation sites, respectively. This result demonstrates that the model in this embodiment can automatically identify pharmacophores and key metabolic structures at the chemical bond level, achieving molecular-level mechanism attribution. This feature interpretation capability enables the model to not only have predictive capabilities but also mechanism identification capabilities, providing a structured and verifiable basis for interaction risk assessment in the early stages of drug development.

[0059] III. Mechanism-level Interpretability Analysis Method Based on Multi-hop Topological Attention To further reveal the structural dependency depth of the model in DDI prediction, this embodiment introduces a multi-hop topological attention mechanism based on Hop2Token and TransformerEncoderReadout to quantify the model's dependence on neighborhood information at different levels. Specifically, as... Figure 6 When the attention weights of higher-order hop numbers (e.g., k=2 or k=3) are significantly higher than those of lower-order hop numbers (k=0 or k=1), it indicates that the interactions predicted by the model depend on complex multi-level metabolic pathways or indirect target associations. For example, the attention distribution of warfarin (DB00946) is concentrated at k=2 (0.280) and k=3 (0.295), revealing its dependence on multi-level regulation of the coagulation pathway; the weights of clopidogrel (DB00374) at k=2 and k=3 are 0.326 and 0.368, respectively, indicating that its mechanism involves metabolic activation processes; in the combination of fluconazole and clozapine (DB01201–DB00214), the model has the highest attention at k=3 (0.371 and 0.366), reflecting the drug accumulation mechanism caused by CYP inhibition.

[0060] This method can reveal potential implicit topological dependencies between drugs, achieving a leap from "observable results" to "explainable mechanisms." Through quantitative multi-hop attention distribution, this invention achieves a precise explanation of drug interaction mechanisms, providing important technical support for drug safety analysis, risk assessment of combined drug use, and traceability of pharmacodynamic pathways.

Claims

1. A method for drug interaction prediction based on multi-modal molecular characterization, characterized in that, The steps are as follows: Step 1: Data set construction Step 2: Building method of feature extraction module The molecular graph topological properties and global semantic information of the drug are uniformly encoded; Step 3: Feature selection and dimension reduction based on SHAP First, a LightGBM model is used to build a drug interaction prediction framework, then the SHAP contribution value of each feature dimension is calculated, and the absolute value mean is sorted in descending order; finally, the top 300 key features are selected for subsequent modeling; Step 4: Building method of drug interaction prediction model 4.1 Graph Attention Network (GAT) First, a graph attention network (GAT) with edge attribute perception is used to update node features; Unlike the standard GCN, which uses uniform or static weighting of neighbors, GAT dynamically calculates attention weights for each neighbor through a self-attention mechanism; let... , It is a node and nodes Features; It is a linear mapping weight matrix of node features, used to calculate the node. and nodes Attention coefficient between As shown in equation (1): (1) wherein, is a learnable attention vector, denotes vector concatenation, is an activation function that normalizes the attention coefficients with Softmax to obtain attention weights as shown in equation (2): (2) In the formula, is a set of adjacent nodes of the node The features of adjacent nodes are aggregated according to attention weights to obtain the node The updated expression is as shown in formula (3): (3) wherein : a non-linear activation function ReLU, : a node an updated representation of the current layer; 4.2 Edge-aware GCNII The first Edge-aware GCNII convolution module, which preserves the initial residual and identity mapping in GCNII, and introduces a learnable attention weight based on edge features, so that the neighborhood aggregation process has distinguishability for different types of edges; layer can be formalized as shown in equation (4): (4) Among them, GCNII is a smooth representation. With the first layer Combine the initial residual join with an added unit mapping. To the Layer weight matrix This approach aims to suppress excessive smoothing in deep networks while simultaneously improving the stability and trainability of parameter updates. 4.3 Hop2Token&Transformer The attention weight of each edge is obtained by linearly mapping the edge features to a scalar As shown in equation (5): (5) wherein, is a learnable weight vector, denotes a set of neighbor nodes of a node including a self-loop node; an attention coefficient together with an edge weight for the neighborhood set; Adjacency matrix Adding an identity matrix And symmetric normalization, get edge-aware normalized adjacency matrix As shown in formula (6), formula (7), formula (8); wherein D is the degree matrix, node The neighborhood aggregation representation in the lth layer is as shown in formula (9): (6) (7) (8) (9) By residual coefficients The original node features and the aggregated features are weighted and fused to obtain intermediate features As shown in equation (10); and then through an identity mapping coefficient The intermediate features are fused with the initial features to obtain the first layer output node features As shown in equation (11): (10) (11) Finally, output to the target dimension through linear mapping, as shown in equation (12): (12) wherein, is a learnable linear mapping matrix, is a bias vector, is the target output dimension; The Hop2Token serialization mechanism is used to convert multi-hop neighborhood information into equivalent Token sequences, which are then input into the Transformer encoder for global dependency modeling; The feature sequence is then input into the Transformer encoder to further model the dependency between hops; 0-hop Token represents the node's own features, i.e. ; k-hop Token represents the k-th order neighbor features, obtained by adjacency matrix multiplication, i.e. k , , ; this process is iterated (i.e. K times); finally, the module stacks all the aggregated features of different hops to generate a multi-scale feature sequence for each node in the graph: (13)​ 4.4 Intra-graph and inter-graph feature extraction input sequence first mapped to query , key and value subspaces, and then attention weights are computed by scaled dot-product attention; its computational form is shown in equation (14): (14) wherein , , , ; In the multi-head mechanism, features are divided into multiple subspaces and calculated in parallel, and finally spliced to form the output, so that each token can achieve weighted aggregation based on the global context of all hop tokens; then, the updated token representation is obtained through LayerNorm and MHSA, as shown in equation (15); the feedforward network independently performs two linear transformations on each token and adds an activation function to improve the non-linear modeling ability of the model, and the output sequence is as shown in equation (16); (15) (16) To obtain the final vector representation of a node, the sequence output by the Transformer encoder is passed through an attention pooling layer that aggregates the token sequences representing neighborhoods of different hop counts into a single node representation that is rich in multi-scale structural information , as shown in equations (17) and (18), where are trainable parameters, and the contribution of each token is determined by the attention ​ (17) (18) 4.5 Adaptive Subgraph Pooling (SAGPooling) structure information compression module Each atom aggregates information and updates its features based on the information sent by its neighbors and its own characteristics. (19) In formula (19), denotes and denotes the attention weight between nodes, and denotes the feature vector of the node and denotes the feature vector of the node and denotes the learnable weight matrix for calculating the attention score; (20) In formula (20) representing the drug in the node after the layer GAT represents, representing the node the set of neighbor nodes of the node, is a normalization constant; for feature extraction of inter-graph structure, the core idea of intra-graph feature extraction is also followed; (21) (22) As shown in equations (21), (22), the inter-graph adjacency matrix is represented as a bipartite graph based on the two drugs ; An adaptive subgraph pooling (SAGPooling) structure information compression method based on attention mechanism is used; The principle of SAGPooling module is shown in equations (19) and (20): 4.6 Multi-modal feature fusion and classification prediction module (23) (24) where, is an adjacency matrix, is the embedding of all nodes in layer, is the trainable matrix of the layer, is the input dimension of the node-level features; Step 5: Model training and validation testing. In each Block, the final drug graph representation consists of three parts, as shown in equation (25): (25) where, from node pooling, from edge feature pooling, from GCNII and Transformer fusion path; subsequently, the graph representation and global molecular features are fused and the final drug representation of Block is obtained by normalization: (26) The results of all Blocks are stacked into a three-dimensional tensor: where B is , L is the number of Blocks, D is the latent space dimension, which is input into the Co-Attention module to perform cross-graph attention interaction between two drugs to obtain cross-drug, and finally the RESCAL knowledge graph embedding module (KGE) combines the relationship vector to score through a bilinear form to realize the interaction prediction of drug pairs, as shown in equation (27): (27) Let the positive sample score be... Negative sample scoring The temperature coefficient is dynamically adjusted based on the standard deviation of the sample scores. As shown in equation (28); The negative samples are given adaptive weights, as shown in equation (29); (28) (29) wherein, is a base temperature coefficient, is an adaptive adjustment coefficient, denotes the sample score standard deviation; the positive and negative sample loss terms are defined as: (30) In formula (30), the boundary offset is , used to expand the discrimination interval of positive and negative samples, is a function, , which represents taking the average of these samples, the positive and negative sample weight coefficients are used to adjust the contribution of different samples in the loss, and the balance coefficient is , used to adjust the loss weight, and the model loss function is defined as follows: (31); The fifth step is specifically as follows:

2. The multi-modal molecular representation based drug interaction prediction method of claim 1, wherein, Five core evaluation indicators are selected for comprehensive evaluation, including accuracy, area under the receiver operating characteristic curve, average precision, recall rate and F1 score; by comparing with current various mainstream baseline models, the advantages of the model in multi-dimensional performance are verified; the experiment design covers transduction tasks and induction tasks, which are used to evaluate the prediction and generalization ability of the model in known drug pairs and new drug pairs. It also includes Step 6, interpretable analysis:

3. The multi-modal molecular representation based drug interaction prediction method of claim 1, wherein, By introducing edge-level attention analysis mechanism, multi-hop topological attention mechanism and atomic-level feature contribution evaluation method into the model structure, multi-level explanation of the drug interaction prediction process is realized; including data input module, graph neural network prediction module and interpretable analysis module, wherein the interpretable analysis module is composed of edge-level attention submodule, multi-hop topological attention submodule and atomic-level contribution analysis submodule; The atomic-level contribution analysis submodule calculates the comprehensive contribution score of each atom in the prediction process through attention pooling and weighted aggregation method, identifies the key atoms and their neighborhood functional fragments, and displays the focus area of the model through visualization technology, realizing the visualization and traceability of the model decision-making process; ​ The edge-level attention sub-module is used for calculating the attention weight of each chemical bond in a molecular structure, and mapping high weight edges to key pharmacophores and potential metabolic sites, so as to reveal the structure-function relationship relied on by the model prediction, and realize the mechanism-consistent structural interpretability; the multi-hop topological attention sub-module generates multi-hop (k-hop) attention weight based on the cooperation of Hop2TokenEncoder and TransformerEncoderReadout, to quantify the degree of dependence of the model on different topological level information when predicting drug interactions, so as to reveal the potential action path of drugs across targets, metabolism and pharmacodynamics.

4. The multi-modal molecular representation based drug interaction prediction method of claim 1, wherein, The first step is specifically operated as follows: 1.1 Data collection and positive and negative sample construction Collect two public DDI benchmark datasets of DrugBank and ChChMiner; the DrugBank dataset is derived from the drug labels of FDA and Health Canada, and contains 1710 drugs and 192,284 interactions, covering 86 interaction types; the negative sample generation method based on Bernoulli distribution is used to generate equal negative samples to realize the balance of positive and negative samples; the ChChMiner dataset contains 998 drugs and 33,669 interaction records, of which 21,486 are positive samples, 12,183 are negative samples, 798 are known drugs, 199 are unknown drugs, and the interaction type is binary classification; 1.2 Data division and evaluation task The conduction task and induction task are designed on the two datasets of DrugBank and ChChMiner respectively; in the conduction task, the DrugBank dataset adopts 3-fold cross-validation, and the training set, validation set and test set are 3:1:1, and ChChMiner adopts 3:1:1 ratio to divide into training set, validation set and test set; in the induction task, in order to simulate the new drug prediction scene, the drugs of the two datasets are randomly divided into new drugs (20%) and old drugs (80%), and the training set only contains old drugs.

5. The multi-modal molecular representation based drug interaction prediction method of claim 1, wherein, The second step is specifically operated as follows: 2.1 Graph structure feature coding module The graph structure coding module is based on the RDKit tool to parse the SMILES string of the drug and generate a molecular graph structure representation; the molecular graph includes an atom set V and an edge set E; the chemical features of each atom and edge are defined according to the chemical field knowledge system, and are linearly mapped and nonlinearly activated through a learnable parameter matrix, thereby forming initial atom feature embedding and edge feature embedding; the non-linear enhancement of the features is realized by introducing the ReLU activation function, so that the topological structure features of the drug molecule at the atomic level can be fully expressed; 2.2 Global molecular descriptor construction module A global molecular descriptor generation mechanism is introduced, which comprehensively uses Mol2Vec and MolT5 two unsupervised molecular representation technologies. The Mol2Vec module maps the molecular substructure to the continuous vector space by analogy with the word embedding idea of natural language, thereby forming a global representation reflecting the similarity of chemical function and structure. The MolT5 module extracts semantic context information from the molecular SMILES sequence based on the multi-layer attention mechanism of the Transformer architecture, generating deep molecular embeddings with high-dimensional semantic features.

Citation Information

Cited By

  • Drug performance prediction method and system based on neural network

    CN122091272A

  • A Drug Interaction Prediction Method and System Based on Knowledge Distillation

    CN122314170A